
Infoscience: Support and Help
- Some of the metrics are blocked by yourconsent settings
Dataset or other product HBB2OBB: Horizontal to Oriented Bounding Box Conversion and Evaluation Tool
(Zenodo, 2026-11-08)HBB2OBB converts horizontal (axis-aligned) bounding boxes (HBBs) into oriented (rotated) bounding boxes (OBBs) by using existing HBB annotations as prompts for segmentation models from the SAM (Segment Anything Model) family. It targets object detection tasks where objects appear at arbitrary orientations, such as aerial imagery, satellite data, and traffic monitoring, producing oriented boxes that tightly encapsulate non-upright objects without any re-annotation. Beyond conversion, it provides horizontal box detection, evaluation, hyperparameter optimization, an interactive annotation viewer, and annotation format-conversion tools, exposed through both a command-line interface and a Python API.
📖 Documentation: rfonod.github.io/hbb2obb, with a quick start on the bundled sample data, a guide per command, pages on how the conversion works, the confidence score, the metrics and provenance, and a CLI and API reference generated from the code.
Why HBB2OBB
- 🎯 Accurate OBBs from HBBs: prompts SAM-family segmentation models with existing horizontal boxes to fit tight oriented boxes around non-upright objects, with no re-annotation required.
- 🚗 No HBBs? Detect them: the
hbb2obb-detectcommand runs an Ultralytics detector over the images and writes the horizontal boxes the conversion consumes, confidence column included. - 🧩 Model ensemble: combines masks from multiple SAM variants through majority voting for more robust, accurate results.
- 🛡️ Spatially constrained and safe: region-specific masking and contour refinement keep segmentation inside the object, a mask split by an occluder is fitted back together, and a fallback retains the original HBB when no valid mask is found, counted and reported rather than left to be discovered afterwards.
- 🔎 Confidence-scored output: every OBB receives a heuristic quality score in [0, 1] that flags silent fallbacks and low-confidence conversions, so you know which boxes to trust; the scores go in the label files or in a side-car directory that keeps them strictly standard, and the detector's own confidence can be carried through instead of, or on top of, that score.
- 📐 Flexible scaling: positive or negative scale factors, optionally different for the short and long sides, recover cropped object parts or tighten overly conservative annotations.
- 📊 Evaluate and optimize: evaluation against ground truth on IoU, orientation error, and the share of boxes above a high IoU bar, plus the
hbb2obb-optimizecommand, a hyperparameter search over SAM inference resolution, scale factors, and morphological opening kernel, driven by a configuration file so that a whole benchmark of model ensembles is a single reproducible command. - 🔬 Where the score comes from: the
hbb2obb-analyzecommand breaks one evaluation down by the ground truth's orientation, size, frame edge, difficulty and class, and scores the conversion against emitting the horizontal prompt unchanged. - 🔄 Five annotation formats: read and write YOLO, DOTA, Pascal VOC, COCO, and LabelMe, for horizontal and oriented boxes alike, with a check that proves every format encodes the same boxes.
- 🔍 Interactive viewer: the
hbb2obb-viewcommand pans and zooms over the annotations, in any format, coloring boxes by confidence and overlaying predictions against ground truth. Where a set ships the same boxes several ways, it names the format on screen and switches between them on a keypress.
Features
- HBB to OBB conversion of YOLO-format annotations using SAM-based segmentation, written in absolute pixels or, with
--normalize, relative to [0, 1], the convention Ultralytics reads. - HBB detection: produce the horizontal boxes in the first place with any Ultralytics detector, a local file, a link, the Ultralytics catalogue, or Hugging Face, passing any other predictor argument straight through.
- Multiple model support: SAM, SAM2, SAM2.1, SAM3, Mobile SAM, and FastSAM families.
- Model ensemble via majority voting for enhanced accuracy.
- Confidence scoring: a per-OBB quality score flags silent fallbacks and low-confidence conversions for triage, optionally combined with the detector confidence from the input, written either as an extra column or to a side-car directory beside standard label files.
- Polygon output: optionally save the segmentation contour behind each OBB, row-aligned with the OBB file, as a tighter object outline for downstream masking.
- Evaluation tools to assess OBB accuracy against ground truth on mean and median IoU, the orientation error of every matched pair, and the share of matched boxes above IoU 0.5, 0.75, 0.85 and 0.9, reported overall and per class. Mean IoU saturates on well-fitted boxes; the other two do not, and separate configurations that it cannot.
- Accuracy breakdown: cut one evaluation by ground-truth orientation, size, frame edge,
difficultflag and class, beside an identity baseline and the converted-to-reference side ratios, written as a report, a YAML and a figure. - Hyperparameter optimization over SAM inference resolutions, HBB scale factors, and morphological opening kernels, one sweep at a time or a whole benchmark from a configuration file, with an aggregate summary and resumable long runs. Figures can be redrawn against any recorded metric without re-running the search, and the benchmark is also written as a LaTeX table and a print-ready PDF figure for a manuscript.
- Provenance records: write the reproducing command, the library versions, the SHA-256 of every checkpoint that ran and how many boxes fell back to their horizontal prompt, beside the annotations, so that a released set can be regenerated rather than trusted. Detections also record the licence and references the Hugging Face Hub declares for their weights, so the credit travels with the labels.
- Visualization tools for HBBs, segmentation masks, derived contours, and resulting OBBs.
- Interactive viewer: pan and zoom over annotated frames, toggle each layer, switch between the formats a set ships, and compare two annotation sets against each other.
- Format conversion utilities between YOLO, DOTA, Pascal VOC, COCO, and LabelMe annotations, in either direction, for both box kinds. A COCO file derived from another can keep the image and annotation identifiers of its source, so the two still join.
Planned Enhancements
- Improved morphological operations for better mask refinement.
- Support for additional segmentation models beyond the SAM and FastSAM families.
Related Projects
- Geo-trax 🚀: georeferenced vehicle trajectory extraction pipeline for high-altitude drone imagery. Its vehicle detector supplies the HBB inputs for vehicle use cases (car, bus, truck, and motorcycle), and
hbb2obb-detect --model geotraxruns that detector directly. - Stabilo ⚖️: Python library for video and trajectory stabilization using robust homography transformations, supporting various feature detectors, RANSAC algorithms, and user-defined masks.
- Stabilo Benchmark 🎯: benchmark of reference-based UAV video stabilization on dynamic urban scenes, accompanying the paper that describes HBB2OBB (arXiv:2611.17190), and the successor of Stabilo-Optimize. HBB2OBB writes the horizontal and oriented boxes behind its object-masking axis (data record).
Citation
If you use HBB2OBB in your research, please cite:
-
Preferred Citation: The paper that describes HBB2OBB:
@article{fonod2026benchmark, author = {Fonod, Robert and Geroliminis, Nikolas}, title = {Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes}, journal = {arXiv preprint arXiv:2611.17190}, year = {2026}, doi = {10.48550/arXiv.2611.17190} } -
Repository Citation: When referencing or building on the code itself (replace
1.x.xwith the version you used):@software{fonod2026hbb2obb, author = {Fonod, Robert}, title = {HBB2OBB: Horizontal to Oriented Bounding Box Conversion and Evaluation Tool}, year = {2026}, month = nov, version = {1.x.x}, license = {MIT}, doi = {10.5281/zenodo.15151143}, url = {https://github.com/rfonod/hbb2obb} }
Contributions
Community contributions are welcome: open a GitHub Issue or submit a pull request.
License
This project is distributed under the MIT License. See the LICENSE file for more details.
- Some of the metrics are blocked by yourconsent settings
Dataset or other product Geo-trax: A Comprehensive Framework for Georeferenced Vehicle Trajectory Extraction from Drone Imagery
(Zenodo, 2026-11-08)Geo-trax (GEO-referenced TRAjectory eXtraction) is a comprehensive pipeline that extracts high-accuracy, georeferenced vehicle trajectories from high-altitude drone imagery. Built for quasi-stationary aerial monitoring of urban traffic, it turns raw bird's-eye view (BEV) drone footage into precise, real-world vehicle trajectories. The framework combines YOLO detection, multi-object tracking, and video stabilization with a robust orthophoto-based georeferencing stage, producing geo-coordinated, lane-resolved trajectories that are spatially and temporally consistent and ready for large-scale traffic analysis and simulation. It is optimized for urban intersections and arterial corridors, where high-fidelity, vehicle-level insights drive intelligent transportation systems and digital twin applications.
📌 Important: If you use this code in your work, kindly acknowledge it by citing the following article:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205
📖 Documentation: rfonod.github.io/geo-trax, with a guide per pipeline stage, setup guides for configuration and your own data, and a CLI and configuration reference generated from the code.
Why Geo-trax
- 🛰️ Real-world output: georeferenced, lane-resolved trajectories (WGS84 + local CRS) with per-vehicle speed, acceleration, and estimated dimensions, straight from raw BEV drone video.
- 🎯 Accurate detection: YOLOv8s vehicle detector reaching 0.951 mAP@50, trained on more than 19,000 annotated aerial images.
- 🚗 Flexible tracking: four vehicle classes and six selectable multi-object trackers (BoT-SORT, ByteTrack, OC-SORT, and more).
- ⚖️ Drone-motion robust: homography-based stabilization (Stabilo) plus orthophoto image registration for consistent, cross-flight coordinates; both optionally CUDA-accelerated.
- 📊 Proven at scale: powered the Songdo Traffic dataset (roughly 700,000 trajectories across 20 intersections, fleet of 10 drones; see Real-World Deployment below).
- ⚙️ One command, one config:
geotrax batchruns the whole pipeline; a single YAML drives every stage, with four tuned presets included.
Features
- Detection: YOLOv8s on aerial BEV imagery; detects car (incl. vans), bus, truck, and motorcycle; optional per-class confidence thresholds (extraction.class_conf) and SAHI sliced inference for improved small-object recall (--sahi).
- Tracking: six multi-object trackers (BoT-SORT default; see Tracking); optional per-track frame-gap interpolation.
- Stabilization: homography-based trajectory correction via Stabilo ⚖️, tuned with Stabilo-Optimize 🎛️; optional CUDA acceleration (--stab-gpu).
- Georeferencing: frame-to-orthophoto registration; outputs lat/lon, local CRS, speed, acceleration, and lane assignment per vehicle; optional CUDA acceleration (--geo-gpu).
- Visualization: track overlays on original, stabilized, or static-reference video, in five rendering modes (incl. oriented bounding boxes).
- Analysis: trajectory maps, kinematic distributions, and class/dimension charts, per-video or aggregated across drones and sessions; GIS export (GeoPackage/GeoJSON) of trajectories for QGIS, ArcGIS, or kepler.gl (
geotrax export). - Scaling & tooling: batch-processes directory trees and aggregates multi-drone data; includes standalone utilities for end-to-end data preparation, training, evaluation, and validation.
🚀 Planned Enhancements
- Modularized, OOP-based pipeline with custom reference frame support and georeferencing leveraging Stabilo's image-matching backend.
- Simulation and scenario export:
geotrax export --format sumo(and/or OpenDRIVE, CommonRoad) to turn an extracted dataset into a drop-in scenario for traffic simulators and digital twins, reusing the lane and road-section geometry already carried by the orthophoto segmentations. - Live/online mode: ingest RTMP-style drone feeds for real-time georeferenced output, for live monitoring and incident detection rather than offline batches.
- Batch inference and multi-thread processing, plus distributed fan-out (Ray, Dask, or Slurm) so
geotrax batchscales across a cluster for multi-drone, multi-intersection campaigns. - Real-world map visualization (e.g., MovingPandas, contextily) and an interactive web viewer with map-based playback and filtering by class, lane, speed, and time, so datasets can be explored without writing plotting code. (
geotrax exportalready produces GeoPackage/GeoJSON layers that open directly in QGIS or kepler.gl.)
🔗 Related Projects
Geo-trax integrates with and complements several specialized tools:
- Stabilo ⚖️: Python library for video and trajectory stabilization using robust homography transformations. Supports various feature detectors, RANSAC algorithms, and user-defined masks. Used as Geo-trax's core stabilization engine.
- Stabilo-Optimize 🎛️: benchmarking and hyperparameter optimization framework for Stabilo. Evaluates stabilization performance through ground truth-free assessment using random perturbations. Used to fine-tune Geo-trax stabilization parameters.
- Stabilo Benchmark 🎯: static and dynamic benchmark of the complete registration pipeline behind reference-based UAV video stabilization (feature detector, matcher, robust estimator and object masking) on urban scenes where vehicles move, from the paper Consensus Capture in Reference-Based UAV Video Stabilization. Evaluates the registration step Geo-trax's stabilization relies on, and annotates its vehicles with the Geo-trax detector.
- HBB2OBB 🔄: converts horizontal bounding boxes to oriented bounding boxes using SAM segmentation models. Can enhance Geo-trax outputs when object orientation is needed for downstream analysis.
Real-World Deployment: The Songdo Experiment
Geo-trax was validated in a large-scale urban traffic monitoring campaign in Songdo, South Korea, where it processed footage from a fleet of 10 drones to produce the Songdo Traffic dataset. The detection model was trained on the companion Songdo Vision dataset. Both are described in the associated publication. Oriented-box annotations of the same frames are released as Songdo Vision OBB, introduced with the stabilization benchmark, which also draws its imagery from this experiment.
Songdo campaign 📍 Location Songdo International Business District, South Korea 📅 Duration 4 days (October 4 to 7, 2022) 🚁 Fleet 10 drones (DJI Mavic 3), 140 to 150 m altitude, 4K at 29.97 fps 🔭 Coverage 20 busy intersections 🚗 Result ~700,000 georeferenced vehicle trajectories 🎥 Demo of Geo-trax applied to the Songdo experiment:
Citation
If you use Geo-trax in your research or software, please cite:
-
Journal article (preferred for any use of the framework):
@article{fonod2025advanced, title = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery}, author = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas}, journal = {Transportation Research Part C: Emerging Technologies}, volume = {178}, pages = {105205}, year = {2025}, publisher = {Elsevier}, doi = {10.1016/j.trc.2025.105205}, url = {https://doi.org/10.1016/j.trc.2025.105205} } -
Software archive (when referencing or building on the code itself; replace
1.x.xwith the version you used):@software{fonod2026geo-trax, author = {Fonod, Robert}, title = {Geo-trax: A Comprehensive Framework for Georeferenced Vehicle Trajectory Extraction from Drone Imagery}, year = {2026}, month = nov, version = {1.x.x}, license = {MIT}, doi = {10.5281/zenodo.12119542}, url = {https://github.com/rfonod/geo-trax} }
Contributions
Early code received key contributions from Haechan Cho (georeferencing) and Sohyeong Kim (video/flight-log merging). Community contributions are welcome: open a GitHub Issue or submit a pull request.
License
This project is distributed under the MIT License. See the LICENSE file for more details.
Full Changelog
- Some of the metrics are blocked by yourconsent settings
Dataset or other product Stabilo Optimize: A Framework for Comprehensive Evaluation and Analysis for the Stabilo Library
(Zenodo, 2026-11-15)> ⚠️ Archived: this repository is archived and no longer developed. Its successor is Stabilo Benchmark 🎯.
Stabilo-Optimize is a Python benchmarking tool designed specifically to evaluate and tune methods and hyperparameters of the Stabilo ⚖️ library for video and track stabilization tasks. It systematically generates performance evaluations through random perturbations, eliminating the need for ground-truth homographies. This tool significantly simplifies the optimization of stabilization techniques, making it ideal for high-precision tasks in fields such as urban monitoring, traffic analysis, and drone imagery processing.
> About the successor: Stabilo Benchmark 🎯 continues this project. Stabilo-Optimize tunes Stabilo under random perturbations of static scenes, the campaign of the associated article. Stabilo Benchmark extends that campaign to complete registration pipelines (feature detector, descriptor, matcher, robust estimator and object masking) on static and real dynamic urban scenes, scored against an audited reference. It is the benchmark of the paper "Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes" (arXiv preprint), whose inputs and results are in a data record on Zenodo.
📌 Important: If you use this code in your work, kindly acknowledge it by citing the following article:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205
Why Stabilo-Optimize
- Ground Truth-Free Benchmarking: Randomly generates photometric and homographic perturbations (brightness variations, Gaussian blur, saturation adjustments, fog effects, rotations, translations, scales, and perspective shifts).
- Hierarchical Benchmarking Strategy: Encourages users to systematically vary hyperparameters hierarchically for efficient parameter optimization.
- Flexible JSON Configuration: Customize extensive parameter settings using nested dictionaries.
- Result Visualization: Generates comprehensive performance plots and benchmarking process visualizations.
Benchmarking Metrics
Benchmarks use metrics like Homography Estimation Accuracy (HEA) and Mean Intersection over Union (MIoU). MIoU specifically evaluates the accuracy of object-level registration and requires bounding box masks for calculation. Detailed metric definitions and analysis are provided in the manuscript.
🔗 Related Projects
- Stabilo ⚖️: the video and trajectory stabilization library that Stabilo-Optimize benchmarks and tunes.
- Stabilo Benchmark 🎯: the successor of Stabilo-Optimize. Extends its static campaign to complete registration pipelines on static and dynamic urban scenes, and is the benchmark of the consensus capture paper, with its data record.
- Geo-trax 🚀: end-to-end framework for extracting georeferenced vehicle trajectories from drone imagery. Uses Stabilo, tuned via Stabilo-Optimize, as its core stabilization engine.
- HBB2OBB 🔄: converts horizontal bounding boxes to oriented bounding boxes using SAM segmentation models. Can be used alongside the Stabilo ecosystem when object orientation is needed for downstream analysis.
Citation
If you use Stabilo-Optimize in your research, software, or product, please cite the following resources appropriately:
-
Preferred Citation: Please cite the associated article for any use of the Stabilo-Optimize, including research, applications, and derivative work:
@article{fonod2025advanced, title = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery}, author = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas}, journal = {Transportation Research Part C: Emerging Technologies}, volume = {178}, pages = {105205}, year = {2025}, publisher = {Elsevier}, doi = {10.1016/j.trc.2025.105205}, url = {https://doi.org/10.1016/j.trc.2025.105205} } -
Repository Citation: When referencing or building on the code itself (replace
1.x.xwith the version you used):@software{fonod2026stabilo-optimize, author = {Fonod, Robert}, title = {Stabilo Optimize: A Framework for Comprehensive Evaluation and Analysis for the Stabilo Library}, year = {2026}, month = nov, version = {1.x.x}, license = {MIT}, doi = {10.5281/zenodo.13828429}, url = {https://github.com/rfonod/stabilo-optimize} } -
Benchmark Citation: If you build on the successor benchmark, or use its software or data, please cite the benchmark paper (see Stabilo Benchmark for its software and data citations):
@article{fonod2026benchmark, author = {Fonod, Robert and Geroliminis, Nikolas}, title = {Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes}, journal = {arXiv preprint arXiv:2611.17190}, year = {2026}, doi = {10.48550/arXiv.2611.17190} }
Contributing
This repository is archived and accepts no new issues or pull requests. For questions and suggestions, please use the issues of Stabilo Benchmark, its successor.
License
This project is distributed under the MIT License. See the LICENSE file for more details.
- Some of the metrics are blocked by yourconsent settings
Dataset or other product Songdo Vision: Vehicle Annotations from High-Altitude BeV Drone Imagery in a Smart City
(Zenodo, 2026-07-31)Overview
The Songdo Vision dataset provides high-resolution 4K RGB images annotated with categorized axis-aligned bounding boxes (BBs) for vehicle detection from a high-altitude bird’s-eye view (BeV) perspective. Captured over Songdo International Business District, South Korea, it comprises 5,419 annotated video frames (4,335 training / 1,084 test; 80/20 split) containing 274,190 vehicle instances across four classes:
- Car (including vans and light-duty vehicles)
- Bus
- Truck
- Motorcycle
This dataset can serve as a benchmark for aerial vehicle detection, supporting research and real-world applications in intelligent transportation systems, traffic monitoring, and aerial vision-based mobility analytics. It was developed in the context of a multi-drone experiment aimed at enhancing geo-referenced vehicle trajectory extraction [1].
A full dataset card with usage instructions and download guidance is available on Hugging Face.
📌 Citation: If you use this dataset in your work, kindly acknowledge it by citing the following article [1]:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205.
🔗 Companion dataset: For precisely georeferenced vehicle trajectories extracted from the same large-scale multi-drone experiment, see Songdo Traffic: 10.5281/zenodo.13828383.
🔗 Oriented bounding boxes: For the same frames with every vehicle annotation of this version converted to an oriented bounding box, see Songdo Vision OBB: 10.5281/zenodo.15050577.
Motivation
Publicly available datasets for aerial vehicle detection often exhibit limitations such as:
- Non-BeV perspectives with varying angles and distortions
- Inconsistent annotation quality, with loose or missing bounding boxes
- Lower-resolution imagery, reducing detection accuracy, particularly for smaller vehicles
- Lack of annotation detail, especially for motorcycles in dense urban scenes with complex backgrounds
To address these challenges, Songdo Vision provides high-quality human-annotated bounding boxes, with machine learning assistance used to enhance efficiency and consistency. This ensures accurate and reliable ground truth for training and evaluating detection models.
Dataset Composition
The dataset is randomly split into training (80%) and test (20%) subsets:
Subset Images Car Bus Truck Motorcycle Total Vehicles Train 4,335 196,719 7,052 11,861 3,030 218,662 Test 1,084 49,862 1,766 3,080 820 55,528 Total 5,419 246,581 8,818 14,941 3,850 274,190 A subset of 5,274 frames was randomly sampled from drone video sequences, while an additional 145 frames were carefully selected to represent challenging cases, such as motorcycles at pedestrian crossings, in bicycle lanes, near traffic light poles, and around other distinctive road markers where they may blend into the urban environment.
Data Collection
The dataset was collected as part of a collaborative multi-drone experiment conducted by KAIST and EPFL in Songdo, South Korea, from October 4–7, 2022.
- A fleet of 10 drones monitored 20 busy intersections, executing advanced flight plans to optimize coverage.
- 4K (3840×2160) RGB video footage was recorded at 29.97 FPS from altitudes of 140–150 meters.
- Each drone flew 10 sessions per day, covering peak morning and afternoon periods.
- The experiment resulted in 12TB of 4K raw video data.
More details on the experimental setup and data processing pipeline are available in [1].
> Note on frame dimensions: 5,411 frames are 3840×2160 (landscape 4K). Eight frames:
05275–05281in the training subset and05282in the test subset are 2160×3840 (portrait 4K), as recorded. Their annotations are correct for that orientation, and the per-image dimensions are stated explicitly in the COCOimagesentries and the Pascal VOC `` element. Normalized YOLO coordinates are relative to each frame's own dimensions. Consumers that assume a fixed 3840×2160 frame size shall read the per-image dimensions instead.Bounding Box Annotations & Formats
Annotations were generated using a semi-automated object detection annotation process in Azure ML Studio, leveraging machine learning-assisted bounding box detection with human verification to ensure precision.
Each annotated frame includes categorized, axis-aligned bounding boxes, stored in three widely-used formats. All three are generated from a single canonical source, so they describe exactly the same boxes: the YOLO files carry sub-pixel precision, while the COCO and Pascal VOC files share one common integer rounding of those coordinates.
-
COCO JSON format
-
Single annotation file per dataset subset (i.e., one for training, one for testing).
-
Contains metadata such as image dimensions, bounding box coordinates, and class labels.
-
Example snippet:
{ "info": {"description": "Songdo Vision: …", "version": "2.0", "year": 2026}, "licenses": [{"id": 1, "name": "Creative Commons Attribution 4.0 International", "url": "https://creativecommons.org/licenses/by/4.0/"}], "images": [{"id": 1, "file_name": "00001.jpg", "width": 3840, "height": 2160}], "annotations": [ {"id": 1, "image_id": 1, "category_id": 2, "bbox": [500, 600, 200, 50], "area": 10000, "iscrowd": 0 } ], "categories": [ {"id": 1, "name": "car", "supercategory": "vehicle"}, {"id": 2, "name": "bus", "supercategory": "vehicle"}, {"id": 3, "name": "truck", "supercategory": "vehicle"}, {"id": 4, "name": "motorcycle", "supercategory": "vehicle"} ] } -
bboxis[x_min, y_min, width, height]in absolute pixels;areaequalswidth × height. -
Image and annotation
ids are globally unique across both subsets. Annotations carried over from v1 keep their original id (all≤ 272435), so records can be traced between releases; annotations added in v2 are numbered from1000001upwards, which makes the two sets easy to tell apart. Ids are not contiguous.
-
-
YOLO TXT format
-
One annotation file per image, following the format:
-
Bounding box values are normalized to [0,1] relative to that frame's own dimensions, with the origin at the top-left corner. Every box lies fully inside the frame.
-
Class ids follow
names.txt:0 = car,1 = bus,2 = truck,3 = motorcycle. -
Example snippet:
0 0.520000 0.630000 0.100000 0.050000 2 0.250000 0.400000 0.150000 0.080000
-
-
Pascal VOC XML format
-
One annotation file per image, structured in XML.
-
Contains image properties and absolute pixel coordinates for each bounding box.
-
`` is set to
1for boxes that touch a frame edge,0otherwise. -
Example snippet:
images 00001.jpg images/00001.jpg <source>Songdo Vision 384021603 0 car Unspecified 0 0 500600600650
-
File Structure
1. Training Data (train.zip, ≈12.9 GB)
train/ │── coco_annotations.json # COCO format │── images/ │ ├── 00001.jpg │ ├── … │── labels/ │ ├── 00001.txt # YOLO format │ ├── 00001.xml # Pascal VOC format │ ├── …2. Testing Data (test.zip, ≈3.2 GB)
test/ │── coco_annotations.json │── images/ │ ├── 00027.jpg │ ├── … │── labels/ │ ├── 00027.txt │ ├── 00027.xml │ ├── …Frame numbering is global across both subsets: an id such as
00027appears in exactly one oftrain/andtest/, never in both.Additional Files
README.md– Dataset documentation (this description)LICENSE.txt– Creative Commons Attribution 4.0 Licensenames.txt– Class names (one per line)data.yaml– Example YOLO configuration file for training/testing
Changes in v2
v2 is an annotation-quality release. The images are byte-identical to v1, only the annotation files,
data.yamlandnames.txtchanged. Ids of surviving v1 annotations are preserved, so v1 and v2 records can be matched directly.In total 196 invalid boxes were removed and 1,951 missed vehicles were added, taking the dataset from 272,435 to 274,190 annotations:
Class v1 Removed Added v2 Car 245,047 140 1,674 246,581 Bus 8,789 10 39 8,818 Truck 14,831 34 144 14,941 Motorcycle 3,768 12 94 3,850 Total 272,435 196 1,951 274,190 Split across subsets, 157 removals and 1,508 additions fall in the training set, and 39 removals and 443 additions in the test set.
Invalid annotations removed
196 of 272,435 boxes (0.072%) were removed. Each geometry threshold was fixed only after visually inspecting the boxes it discards, and every
false_positiveremoval was inspected individually.Reason Removed What it catches false_positive124 No vehicle at all: painted road markings, rooftop equipment, construction machinery, manhole covers side_too_small43 Shorter side below 6 px (smallest genuine motorcycle side is ~12 px) duplicate12 One vehicle annotated twice (IoU ≥ 0.5 and comparable box areas) extreme_aspect_ratio9 Aspect ratio above 6 and shorter side below 35 px implausibly_large5 Longer side above 500 px (largest genuine vehicle is ~403 px) curated2 Single box spanning several diagonally parked cars area_too_small1 Area below 150 px² (1st-percentile motorcycle area is ~520 px²) Total 196 The thresholds are deliberately far outside the legitimate distribution. In particular, the aspect-ratio rule also requires a short side below 35 px, because genuine articulated trucks in this dataset reach an aspect ratio of ~6.9 and would otherwise have been discarded.
Detector-assisted review
v1 was produced by a semi-automated annotation pass with human verification. That process is good at precision but can still miss objects, so v2 re-examined the annotations against an independent YOLOv8s vehicle detector trained on high-altitude BeV drone imagery (geo-trax). The detector was used only to rank candidates for human review, it never edited an annotation on its own, and every change below was confirmed by a human looking at the image.
False positives removed. Every annotation was scored against the detector, and the boxes it did not support were inspected individually. Confirmed non-vehicles, road markings and painted text, rooftop equipment, construction machinery, manhole covers, were removed.
Missed vehicles added (1,951 annotations). Confident detections (≥ 0.70) that overlapped no existing annotation were reviewed the same way; confirmed vehicles were added. This signal is stronger than it may appear: the detector was trained on these very annotations, so training actively pushed it to stay silent wherever the ground truth was empty. A confident detection in an unannotated spot is therefore good evidence of a genuine miss rather than a model artifact.
Added boxes take their geometry and class from the detector, with a human confirming that a vehicle of that class is present. They carry COCO ids from
1000001upwards, so anyone who wants only the original v1-provenance annotations can filter them out.Review coverage: all 1,983 addition candidates and all 1,053 false-positive candidates were inspected on contact sheets; every box marked for change was then re-checked zoomed. A random sample of 72 accepted additions was independently re-checked, of which 71 were correct, so a small residual error rate (order 1%) should be assumed for the added boxes.
Consistency and correctness fixes
- All three formats now agree exactly. In v1, COCO truncated the box width while Pascal VOC truncated the box corner, so the two disagreed by up to 1 px on roughly 75% of boxes. Both are now derived from one shared rounding of the same source coordinates.
- 1,464 boxes extended microscopically past the frame border (by at most 0.002 px, a rounding artifact). All boxes are now strictly inside their frame, and every normalized YOLO coordinate lies within
[0,1]. - Duplicate annotations removed. Twelve vehicles carried two overlapping boxes, usually with conflicting class labels. A pair is treated as a duplicate only when the boxes overlap at IoU ≥ 0.5 and have comparable areas (ratio ≥ 0.5), so the axis-aligned box of a diagonally oriented bus can never absorb a small vehicle beside it. The tighter box was kept, except for one pickup truck where the larger box correctly covers both cab and bed.
- COCO gained the standard
infoandlicensesblocks and asupercategoryfield on each category. - Pascal VOC gained
,,andper object, which several standard VOC parsers require. data.yamlnow points at the concept DOI (which always resolves to the latest version) instead of the v1-specific DOI, and itspathfield follows the Ultralytics convention: a relativepathis resolved against the Ultralytics datasets directory, not against the location ofdata.yaml. The v1 value (../datasets/songdo_vision) only worked for one particular layout.names.txtnow ends with a newline.- The 8 portrait frames are documented (see Data Collection); they were undocumented in v1, which described every frame as 3840×2160.
Known limitations
- Class boundaries. Car/van/truck assignments on v1 annotations were not revised. The detector disagreed with the annotated class on roughly 0.55% of boxes, but these are largely genuine judgement calls (e.g. large vans, pickups) that the original annotators owned.
- Remaining missed vehicles. The additions came from detections at confidence ≥ 0.70; vehicles the detector found with lower confidence, or missed entirely, are still unannotated. Recall is improved, not complete.
- Test-set provenance. Because the added test-split boxes were surfaced by a geo-trax detector, benchmarking that same model family on Songdo Vision test carries a small favorable bias. It is bounded, the additions are 0.8% of the test annotations, and every one was human-confirmed, but it is worth stating for strict benchmarking.
- Loosely overlapping pairs. 104 box pairs overlap at an IoU between 0.3 and 0.5, and one box fully contains two others. These are almost always correct, the axis-aligned box of a diagonally parked car naturally encloses its neighbours, so they were flagged rather than changed.
Acknowledgments
In addition to the funding sources listed in the metadata, the creators express their gratitude to Artem Vasilev for his dedicated efforts in data annotation. We also thank the research teams of Prof. Simon Oh (Korea University) and Prof. Minju Park (Hannam University) for their assistance during the data collection campaign, including the provision of drone equipment and student support.
Citation & Attribution
Preferred Citation: If you use Songdo Vision for any purpose, whether academic research, commercial applications, open-source projects, or benchmarking efforts, please cite our accompanying article [1]:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205 (Preprint: arXiv:2411.02136) > > BibTeX entry: > >
> @article{fonod2025advanced, > title = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery}, > author = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas}, > journal = {Transportation Research Part C: Emerging Technologies}, > volume = {178}, > pages = {105205}, > year = {2025}, > publisher = {Elsevier}, > doi = {10.1016/j.trc.2025.105205}, > url = {https://doi.org/10.1016/j.trc.2025.105205} > } >Dataset Citation (for archival purposes): Although Zenodo automatically provides a formal dataset citation (shown below), including citation export in various formats such as BibTeX, we kindly request that you reference the above article as the primary source of this work.
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2026). Songdo Vision: Vehicle Annotations from High-Altitude BeV Drone Imagery in a Smart City (v2). Zenodo. DOI: 10.5281/zenodo.21728430
Version History
-
v1 (Initial release, March 2025)
Initial public release: 5,419 annotated frames with 272,435 vehicle instances in COCO JSON, YOLO TXT and Pascal VOC XML.
-
v2 (Current release, July 2026)
Annotation-quality release on byte-identical images: 196 invalid boxes removed, 1,951 missed vehicles added (274,190 in total), the three annotation formats reconciled to a single canonical source, and
data.yamlandnames.txtcorrected. Ids of surviving v1 annotations are preserved, so records can be matched between releases. See Changes in v2 above.
- Some of the metrics are blocked by yourconsent settings
Dataset or other product Songdo Traffic: High Accuracy Georeferenced Vehicle Trajectories from a Large-Scale Study in a Smart City
(Zenodo, 2025-12-13)Overview
The Songdo Traffic dataset delivers precisely georeferenced vehicle trajectories captured through high-altitude bird's-eye view (BeV) drone footage over Songdo International Business District, South Korea. Comprising approximately 700,000 unique trajectories, this resource represents one of the most extensive aerial traffic datasets publicly available, distinguishing itself through exceptional temporal resolution that captures vehicle movements at 29.97 points per second, enabling unprecedented granularity for advanced urban mobility analysis.
📌 Citation: If you use this dataset in your work, kindly acknowledge it by citing the following article:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205.
🔗 Companion dataset: For high-resolution annotated images with vehicle bounding boxes supporting aerial detection research, see Songdo Vision: 10.5281/zenodo.13828407.
Dataset Composition
The dataset consists of five primary components:
- Trajectory Data: 80 ZIP archives containing high-resolution vehicle trajectories with georeferenced positions, speeds and acceleration profiles, and other metadata.
- Orthophoto Cut-Outs: High-resolution (8000×8000 pixel) orthophoto images for each monitored intersection, together with the georeferencing parameters needed to convert orthophoto pixels into real-world coordinates.
- Road and Lane Segmentations: CSV files defining lane polygons within road sections, facilitating mapping of vehicle positions to road segments and lanes.
- Master Frames: Reference images and transformation parameters relating stabilized video frames to the orthophoto cut-out coordinate system, for users who wish to reconstruct or audit the georeferencing.
- Sample Videos: A selection of 4K UHD drone video samples capturing intersection footage during the experiment.
Data Collection
The dataset was collected as part of a collaborative multi-drone experiment conducted by KAIST and EPFL in Songdo, South Korea, from October 4–7, 2022.
- A fleet of 10 drones monitored 20 busy intersections, executing advanced flight plans to optimize coverage.
- 4K (3840×2160) RGB video footage was recorded at 29.97 FPS from altitudes of 140–150 meters.
- Each drone flew 10 sessions per day, covering peak morning and afternoon periods.
- The experiment resulted in 12TB of 4K raw video data.
More details on the experimental setup and data processing pipeline are available in [1].
Data Processing
The trajectories were extracted using 🚀 Geo-trax, an advanced deep learning framework designed for high-altitude UAV-based traffic monitoring. This state-of-the-art pipeline integrates vehicle detection, tracking, trajectory stabilization, and georeferencing to extract high-accuracy traffic data from drone footage.
🎥 A demonstration of the Geo-trax framework in operation is available at:
Key Processing Steps:
- Vehicle Detection & Tracking: Vehicles were detected and tracked across frames using a deep learning-based detector and motion-model-based tracking algorithm.
- Trajectory Stabilization: A novel track stabilization method was applied using detected vehicle bounding boxes as exclusion masks in image registration.
- Georeferencing & Coordinate Transformation: Each trajectory was transformed into global (WGS84), local Cartesian, and orthophoto coordinate systems.
- Vehicle Metadata Estimation: In addition to time-stamped vehicle trajectories, various metadata attributes were also extracted, including vehicle dimensions and type, speed, acceleration, class, lane number, road section, and visibility status.
More details on the extraction methodology are available in [1].
File Structure & Formats
1. Trajectory Data (Daily Intersection ZIPs, 16.2 MB ~ 360.2 MB)
The trajectory data is organized into 80 ZIP files, each containing traffic data for a specific intersection and day of the experiment.
File Naming Convention:
YYYY-MM-DD_intersectionID.zipYYYY-MM-DDrepresents the date of data collection (2022-10-04 to 2022-10-07).intersectionIDis a unique identifier for one of the 20 intersections where data was collected (A, B, C, E, …, U). The letter D is reserved to denote "Drone".
Each ZIP file contains 10 CSV files, each corresponding to an individual flight session:
YYYY-MM-DD_intersectionID.zip │── YYYY-MM-DD_intersectionID_AM1.csv ├── … │── YYYY-MM-DD_intersectionID_AM5.csv │── YYYY-MM-DD_intersectionID_PM1.csv ├── … └── YYYY-MM-DD_intersectionID_PM5.csvHere, AM1-AM5 and PM1-PM5 denote morning and afternoon flight sessions, respectively. For example,
2022-10-04_S_AM1.csvcontains all extracted trajectories from the first morning session of the first day at the intersection 'S'.CSV File Example Structure:
Each CSV file contains high-frequency trajectory data, formatted as follows (d.p. = decimal place):
Dataset Column Name Format / Units Data Type Explanation Vehicle_ID1, 2, … Integer Unique vehicle identifier within each CSV file Local_Timehh:mm:ss.sssString Local Korean time (GMT+9) in ISO 8601 format Drone_ID1, 2, …, 10 Integer Unique identifier for the drone capturing the data Ortho_X,Ortho_Ypx (1 d.p.) Float Vehicle center coordinates in the 8000 × 8000 px orthophoto cut-out image (see Reproducing Georeferenced Coordinates) Local_X,Local_Ym (2 d.p.) Float KGD2002 / Central Belt 2010 planar coordinates (EPSG:5186) Latitude,Longitude° DD (7 d.p.) Float WGS84 geographic coordinates in decimal degrees (EPSG:4326) Vehicle_Length*,Vehicle_Width*m (2 d.p.) Float Estimated physical dimensions of the vehicle Vehicle_ClassCategorical (0–3) Integer Vehicle type: 0 (car/van), 1 (bus), 2 (truck), 3 (motorcycle) Vehicle_Speed*km/h (1 d.p.) Float Estimated speed computed from trajectory data using Gaussian smoothing Vehicle_Acceleration*m/s² (2 d.p.) Float Estimated acceleration derived from smoothed speed values Road_Section*N_GString Road section identifier (N = node, G = lane group) Lane_Number*1, 2, … Integer Lane position (1 = leftmost lane in the direction of travel) Visibility0/1 Boolean 1 = fully visible, 0 = partially visible in the camera frame * These columns may be empty under certain conditions, see [1] for more details.
2. Orthophoto Cut-Outs & Georeferencing Parameters (orthophotos.zip, 1.8 GB)
For each monitored intersection, we provide high-resolution orthophoto cut-outs derived from a comprehensive, high-accuracy orthophoto mosaic covering the entire experimental area. These 8000 × 8000 pixel PNG images facilitate accurate georeferencing and contextual visualization of vehicle trajectories.
orthophotos/ │── ortho_parameters.txt │── A.png │── A_center.txt │── B.png │── B_center.txt │── … │── U.png └── U_center.txtIncluded Files:
ortho_parameters.txt– Georeferencing parameters of the base (global) orthophoto, defining the mapping between base orthophoto pixel coordinates and real-world coordinates. Contains four values, one per line and in this order: base longitude (lng₀), base latitude (lat₀), longitude per pixel (Δlng), and latitude per pixel (Δlat), all at the base orthophoto's native resolution.*.png– High-resolution orthophoto cut-out for a given intersection.*_center.txt– Per-intersection offset files defining the center of each cut-out in the base orthophoto's pixel coordinate system. Contains two values, one per line: center_x, then center_y.
Reproducing Georeferenced Coordinates from Orthophoto Coordinates
The orthophoto cut-outs provided here (8000 × 8000 px) were downscaled from original 15,000 × 15,000 px cut-outs extracted from the base orthophoto. The
Ortho_XandOrtho_Ycolumns in the trajectory CSVs refer to positions within these 8000 × 8000 px images. However, the center coordinates in*_center.txtand the pixel scale inortho_parameters.txtare defined at the original 15,000 × 15,000 resolution. To correctly convert orthophoto coordinates to geographic (WGS84) coordinates, the following steps are required:- Read the base orthophoto parameters from
ortho_parameters.txt: lng₀, lat₀, Δlng, Δlat. - Read the cut-out center from
{ID}_center.txt: center_x, center_y (in base orthophoto pixels). - Compute the cut-out origin (top-left corner) using the original cut-out size (15,000 px):
- origin_x = center_x − 7500
- origin_y = center_y − 7500
- Compute the adjusted geographic origin:
- lng₀' = lng₀ + origin_x × Δlng
- lat₀' = lat₀ + origin_y × Δlat
- Scale the pixel size to account for the downscaling from 15,000 to 8,000 pixels:
- Δlng' = Δlng × (15000 / 8000)
- Δlat' = Δlat × (15000 / 8000)
- Convert orthophoto coordinates to geographic coordinates:
- Longitude = lng₀' + Ortho_X × Δlng'
- Latitude = lat₀' + Ortho_Y × Δlat'
For more details on the orthophoto generation process and the choice of cut-out resolution, refer to [1].
3. Orthophoto Segmentations (segmentations.zip, 24.9 KB)
We provide the road and lane segmentations for each orthophoto cut-out, stored as CSV files where each row defines a lane polygon within a road section.
Each section (
N_G) groups lanes moving in the same direction, with lanes numbered sequentially from the innermost outward. The CSV files are structured as follows:segmentations/ │── A.csv │── B.csv │── … └── U.csvEach file contains the following columns:
Section: Road section ID (N_Gformat).Lane: Lane number within the section.tlx, tly,blx, bly,brx, bry,trx, try: Polygon corner coordinates (top-left, bottom-left, bottom-right, top-right), in the pixel coordinate system of the 8000 × 8000 px orthophoto cut-out for that intersection, the same space as theOrtho_XandOrtho_Ytrajectory columns.
These segmentations enabled trajectory points to be mapped to specific lanes and sections in our trajectory dataset. Vehicles outside segmented areas (e.g., intersection centers) remain unlabeled. Perspective distortions may also cause misalignments for taller vehicles.
4. Master Frames (master_frames.zip, 248.6 MB)
We include master reference frames and associated transformation parameters used to align stabilized video frames with the orthophoto space. These are helpful for users wishing to reconstruct or audit the georeferencing transformations applied during trajectory extraction.
master_frames/ │── A.png │── A.txt │── B.png │── B.txt │── … │── U.png └── U.txtIncluded Files:
*.png– Master reference image for a given intersection.*.txt– Corresponding text file containing the transformation parameters used to relate the master frame to the orthophoto cut-out coordinate system: the nine coefficients of the 3 × 3 homography matrix in row-major order (the last element normalized to 1) on the first line, followed by the MD5 hash of the master frame and the keypoint-matching statistics of the registration.
5. Sample Videos (sample_videos.zip, 26.8 GB)
The dataset includes 29 video samples, each capturing the first 60 seconds of drone hovering over its designated intersection during the final session (PM5) on October 7, 2022. These high-resolution 4K videos provide additional context for trajectory analysis and visualization, complementing the orthophoto cut-outs and segmentations.
sample_videos/ │── A_D1_2022-10-07_PM5_60s.mp4 │── A_D2_2022-10-07_PM5_60s.mp4 │── B_D1_2022-10-07_PM5_60s.mp4 │── … └── U_D10_2022-10-07_PM5_60s.mp4Additional Files
README.md– Dataset documentation (this description)LICENSE.txt– Creative Commons Attribution 4.0 License
Known Dataset Artifacts and Limitations
While this dataset is designed for high accuracy, users should be aware of the following known artifacts and limitations:
- Trajectory Fragmentation: Trajectories may be fragmented for motorcycles in complex road infrastructure scenarios (pedestrian crossings, bicycle lanes, traffic signals) and for certain underrepresented truck variants. Additional fragmentations occurred when drones experienced technical issues during hovering, necessitating mid-recording splits that naturally resulted in divided trajectories.
- Vehicle ID Ambiguities: The largest
Vehicle_IDin a CSV file does not necessarily indicate the total number of unique vehicles. - Kinematic Estimation Limitations: Speed and acceleration values are derived from raw tracking data and may be affected by minor errors due to detection inaccuracies, stabilization artifacts, and applied interpolation and smoothing techniques.
- Vehicle Dimension Estimation: Estimates may be unreliable for stationary or non-axially moving vehicles and can be affected by bounding box overestimations capturing protruding vehicle parts or shadows.
- Lane and Section Assignment Inaccuracies: Perspective effects may cause vehicles with significant heights, such as trucks or buses, to be misassigned to incorrect lanes or sections in the orthophoto.
- Occasional pedestrian pair misclassifications: Rarely, two pedestrians walking side by side may be briefly mistaken for a motorcycle, but such instances are short-lived and typically removed by the short trajectory filter.
For a comprehensive discussion of dataset limitations and validation procedures, refer to [1].
Acknowledgments
In addition to the funding sources listed in the metadata, the creators express their gratitude to Sohyeong Kim, Yura Tak, and Weijiang Xiong for their invaluable assistance in the data-wrangling process, Artem Vasilev for his dedicated efforts in data annotation, Jasso Espadaler Clapés for his insightful contributions to vehicle dimension estimation, and Muhammad Ahmed for his support in creating the experimental orthophoto for validation purposes and his contributions to road segmentation. We also thank DroMii company for producing the primary orthophoto and Stanford Center at the Incheon Global Campus (SCIGC) for providing the on-board recorded trajectories of their autonomous vehicle. We also thank the research teams of Prof. Simon Oh (Korea University) and Prof. Minju Park (Hannam University) for their assistance during the data collection campaign, including the provision of drone equipment and student support.
Citation & Attribution
Preferred Citation: If you use Songdo Traffic for any purpose, whether academic research, commercial applications, open-source projects, or benchmarking efforts, please cite our accompanying article [1]:
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205 > > BibTeX entry: > >
> @article{fonod2025advanced, > title = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery}, > author = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas}, > journal = {Transportation Research Part C: Emerging Technologies}, > volume = {178}, > pages = {105205}, > year = {2025}, > publisher = {Elsevier}, > doi = {10.1016/j.trc.2025.105205}, > url = {https://doi.org/10.1016/j.trc.2025.105205} > } >Dataset Citation (for archival purposes): Although Zenodo automatically provides a formal citation for this dataset (see below), including citation export in various formats such as BibTeX, we kindly request that you reference the above article as the primary source of this work.
> Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Songdo Traffic: High Accuracy Georeferenced Vehicle Trajectories from a Large-Scale Study in a Smart City (v2). Zenodo. DOI: 10.5281/zenodo.17924857.
Version History
-
v1 (Initial release, March 2025)
Initial public release of the Songdo Traffic dataset, including georeferenced vehicle trajectories, orthophoto cut-outs, road and lane segmentations, and sample videos.
-
v2 (Current release, December 2025)
Added master reference frames used in the georeferencing pipeline and included the previously missing orthophoto georeferencing .txt parameter files, enabling full reproducibility of the coordinate transformations. Updated the dataset description with instructions for reproducing georeferenced coordinates from the provided orthophoto cut-out coordinates, clarifying the relationship between the provided 8000 × 8000 px cut-outs and the original 15,000 × 15,000 px resolution.
- Some of the metrics are blocked by yourconsent settings
Publication Ultra-Miniaturized and Ultra-Low-Power CMOS Interfaces for In-Body Biosensing
(EPFL, 2026)Implantable biomedical systems are transitioning from large, battery-powered single-function devices to networks of microscale, wirelessly powered sensor nodes that monitor the brain's chemical and electrical state from within. Concepts such as Neural Dust and Body Dust envision implant nodes the size of blood cells, harvesting power wirelessly within a nanowatt energy budget. Realizing this vision requires electrochemical interfaces capable of detecting neurotransmitters such as dopamine and glutamate in vivo, since these molecules govern neurological processes invisible to electrical recording alone. Conventional CMOS potentiostats, however, rely on analog transimpedance amplifiers and explicit analog-to-digital converters that typically consume microwatts and occupy tenths of a square millimeter, three or more orders of magnitude beyond what sub-mm implants allow. This thesis addresses that gap.
The work is organized around a fully digital potentiostat architecture in which sensing, digitization, and closed-loop electrode potential control are unified into a single compact block based on a digital operational transconductance amplifier. This design, with co-designed supporting circuits and CMOS-compatible microfabricated probe electrodes, is validated from sensor modeling and individual circuit blocks to a complete system-level neurochemical sensing platform, and is further extended to neural signal acquisition through a miniaturized intravascular recording front-end.
The thesis reports six principal contributions. First, a more reliable equivalent circuit model for electrochemical sensors incorporating a Butlerâ Volmer voltage-dependent current source and constant-phase elements at both electrode interfaces, addressing CMOS simulation errors of up to 60% in current and 110% in saturation-risk prediction. Second, the fully digital potentiostat, fabricated in TSMC 180 nm CMOS at 0.4 V, consuming only 1.2 nW in 260 µm2 (excluding reference and oscillator), the lowest power and smallest area ever reported to date. Third, a capacitor-less Dynamic Leakage Suppression ring oscillator achieving 69 µm and 7.9 nW at 21.89 kHz. Fourth, two complementary low-dropout regulators delivering 400 mV at quiescent currents of 7.3 and 21.5 nA in 1563.7 and 338.8 µm2, respectively. Fifth, CMOS-compatible foldable platinum microprobes integrated with the digital potentiostat to form a complete platform, achieving dopamine detection at 1383 pulses/µM sensitivity and 151 nM limit of detection, and enzymatic glutamate detection at 269 pulses/µM and 1.2 µM limit of detection. Sixth, a 0.0025 mm2 intravascular neural front-end combining a differential capacitive-feedback low-noise amplifier with a VCO-based frequency encoder, achieving 5.55 µVrms input-referred noise over the spike band at 11 µW for catheter-deployable recording with body-coupled communication.
Together, these contributions demonstrate that neurochemical sensing can be performed by a fully integrated CMOS system within the nanowatt power and sub-mm2 area envelope of distributed implantable devices. The results establish a validated design methodology spanning sensor co-design, digital circuit architecture, power management, clock generation, microfabrication, and system integration, providing a reusable foundation for next-generation implantable biosensors and bringing the Body Dust vision closer to clinical realization.
12 - Some of the metrics are blocked by yourconsent settings
Publication Dimension-Reduced Surrogates and Memory-Assisted Projections for Expensive Parametric and Evolution Problems
(EPFL, 2026)Many scientific computing tasks repeatedly evaluate expensive models to quantify uncertainty, optimize a design, or approximate the solution of a dynamical system. Yet successive computations often share structure because they revisit either the same target map or nearby ones. This thesis investigates how information that has already been acquired can reduce the cost of the computations that follow. It is organized around two questions:
- (1) When the target is fixed, how can expensive information be used to construct a reusable surrogate for later queries?
- (2) When the target evolves along a trajectory, how can information from previous steps predict the approximation needed at the current one, so that only limited new expensive information must be computed?
The first part of the thesis addresses fixed target maps by constructing global surrogates that can be evaluated repeatedly without calling the expensive model. We build on the active-subspace method, which identifies a few linear combinations of the input parameters along which the model output varies most. We prove that coordinates learned from finitely many gradient evaluations of the target map are, with high probability, nearly as accurate as the best reduced coordinates in the error measure relevant to the resulting surrogate. The number of required gradients is controlled by how concentrated the informative parameter directions are under the sampling distribution. Known smoothness of the parametric model then provides practical sampling rules.
We then propose a multilevel version of active-subspace surrogate construction, assigning separate reduced coordinates and polynomial approximations to the corrections between successive model fidelities. This allocation uses many evaluations of inexpensive coarse models to capture the dominant behavior and reserves costly fine-model evaluations for the smaller details that remain. For an idealized construction, the analysis quantifies when this strategy reduces work compared with a single-level method. An adaptive implementation produces an evaluable surrogate, and numerical experiments demonstrate lower construction work at comparable accuracy.
The second part addresses targets that evolve during optimization or time integration. For stochastic optimization, we propose a method that fits stored sample gradients by weighted least squares to predict the current objective gradient and uses a fresh sample-gradient difference to correct the prediction without introducing bias. The number of fresh gradients remains controlled throughout the iterations, thus past gradient evaluations replace most of the expensive new data that would otherwise be needed at each design.
For the approximation of nonlinear dynamical systems we build on the Dirac-Frenkel variational principle, which evolves a parametrized state by fitting the system's vector field in the tangent space induced by the parametrization. We propose an inertial extension that retains the previous parameter velocity and solves a regularized tangent least-squares problem only for its correction. We establish well-posedness and a posteriori state-error estimates in continuous and discrete time. When consecutive velocities are sufficiently coherent, the method can use stronger regularization and smaller randomized sketches while retaining accuracy, thereby reducing the cost of the correction computed at each time step. This behavior is supported by numerical experiments.
11 - Some of the metrics are blocked by yourconsent settings
Publication Multipanel Point-of-Care Electrochemical Biosensing Platform for Therapeutic Drug Monitoring in Precision Oncology
(EPFL, 2026)Chemotherapy remains a cornerstone of cancer treatment, yet its clinical management is complicated by narrow therapeutic windows, pronounced inter-individual pharmacokinetic variability, and the routine co-administration of multiple drugs. These factors create a strong clinical need for therapeutic drug monitoring (TDM), which adjusts dosing to keep each patient within the therapeutic range. Current TDM relies on chromatographic and mass-spectrometric methods that are accurate but slow, costly, and confined to centralized laboratories, preventing the real-time, bedside feedback required for individualized dose optimization in oncology.
This thesis develops an electrochemical biosensing platform for point-of-care TDM of six clinically relevant chemotherapeutic agents — cyclophosphamide, ifosfamide, etoposide, methotrexate, 5-fluorouracil, and busulfan — selected for their narrow therapeutic indices and high inter-individual variability. Because these drugs present a wide range of electrochemical behaviors, two complementary detection strategies were developed. Intrinsically electroactive drugs (etoposide, methotrexate, 5-fluorouracil) were detected directly at carbon electrodes; their simultaneous detection on a single sensor was demonstrated, and a previously uncharacterized mutual electrochemical cross-reactivity between etoposide and methotrexate was identified and mechanistically explained. Carboxylated multi-walled carbon nanotubes were further shown to selectively amplify the first oxidation peak of etoposide through an adsorption-mediated interfacial process.
Non-electroactive drugs were addressed through enzyme-mediated transduction. Cytochrome P450 biosensors were built on nanostructured supports, comparing spinel oxide nanoparticles with platinum nanopetals; the latter enabled, to the best of our knowledge, the first demonstration of mediator-free direct electron transfer with CYP2B6, yielding a sensitivity exceeding previously reported CYP-based biosensors by more than one order of magnitude. For busulfan, which is neither electroactive nor a CYP substrate, the first inhibition-based electrochemical biosensor was realized through a glutathione S-transferase mechanism.
These strategies were integrated into a dual-channel biosensor combining direct and enzymatic detection on a single substrate, enabling the detection of multiple drugs and their simultaneous measurement in clinically relevant combinations, without cross-interference between channels. The platform was validated in undiluted human serum, where matrix effects and electrode biofouling were mitigated using drug-specific strategies, including acetonitrile protein precipitation and a protective chitosan layer for the enzymatic electrodes. Finally, a dedicated, battery-powered electrochemical front-end supporting cyclic and differential pulse voltammetry, therapy-based operation, and wireless communication was designed and integrated onto a compact printed circuit board, reproducing benchtop-instrument performance in a portable form.
Together, these contributions deliver a complete proof of concept for portable, multi-drug electrochemical TDM, a concrete step toward closed-loop, individualized chemotherapy management at the patient's bedside.
15 - Some of the metrics are blocked by yourconsent settings
Publication Bayesian Plasma Tomography in Tokamak Fusion Devices: A Unifying Framework, Real-Time Inference, and Learned Priors
(EPFL, 2026)Nuclear fusion promises a low-carbon and potentially near-limitless energy source. Currently, the leading approach to fusion power is the magnetic confinement of high-temperature ionized gases called plasmas, in devices such as tokamaks and stellarators. Present-day experiments aim to understand and solve the physics and engineering problems that stand in the way, relying on diagnostics that probe the plasma's properties and performance. Among these are systems measuring the radiation emitted by the plasma in different energy ranges, such as soft X-ray detectors or broadband bolometers. To a first approximation, these detectors measure the plasma radiation integrated along their lines-of-sight. Tomography then reconstructs the spatial emissivity distribution from these data. Technical and design constraints, however, drastically limit the number of measurements. Estimating the emissivity profiles is thus a severely underdetermined problem: for a given set of measurements, infinitely many profiles are perfectly consistent with the data.
Bayesian statistics provides a systematic and principled way to account for the resulting uncertainty. Statistical modeling of the observed data leads to a likelihood term, while the assumed properties of the profile are encoded in a prior term; together, they yield the posterior, which models all the available information about the profile. The prior compensates for the missing information, at the price of biasing the solution.
In this thesis, we develop a unifying Bayesian framework for plasma tomography, framing the problem in the language of modern computational imaging. Using robust optimization and sampling algorithms, we show how to extract uncertainty-aware estimates of quantities of interest from the posterior; the most popular tomography techniques are recovered as deterministic point estimates within this framework. We then compute, for the TCV tokamak at EPFL, physically accurate volume-of-sight models of the diagnostic views; these improve the overall reconstruction while barely affecting the estimated power radiated by large plasma regions, already well captured by simpler line-of-sight models. We propose a technique providing accurate real-time estimates of the radiated power as an analytically derived linear combination of bolometer measurements. Its coefficients are pre-computed before a given discharge, allowing scenario-dependent adaptation and tracking radiation from arbitrary regions of interest. We report on the first TCV experiments to rely on real-time bolometry, where plasma exhaust was actively controlled with our technique. Finally, we investigate data-driven priors based on diffusion models, learned from physically realistic profiles. The resulting model, despite not relying on the magnetic equilibrium, can outperform hand-designed equilibrium-informed priors, and its estimated uncertainty better localizes the reconstruction error.
All the developed approaches are facets of the same general Bayesian framework. Given the extremely underdetermined nature of plasma tomography, none of them can be expected to be flawless: robustness, accuracy, speed, interpretability and model expressiveness cannot be maximized simultaneously. We explore these fundamental trade-offs and argue that a sound treatment of this complex problem should combine different probabilistic methods, benefiting from the best features of each. All the developed computational routines are made openly available.
- Some of the metrics are blocked by yourconsent settings
Publication Construire l'ordre international des brevets : l'Union de Paris et la gouvernance des divergences (Allemagne, États-Unis, Royaume-Uni, 1883-1958)
(EPFL, 2026)Cette thèse retrace la construction et les transformations de l’ordre international des brevets autour de l’Union de Paris, de sa fondation en 1883 à l’après-Seconde Guerre mondiale, principalement à partir des cas allemand, états-unien, français et britannique. Elle interroge la manière dont le brevetage multinational a été concilié avec la persistance de systèmes nationaux profondément différents et la volonté des États de préserver leur souveraineté juridique, économique et technologique.
À rebours d’une histoire conçue comme une progression vers l’unification du droit des brevets, cette recherche montre que la Convention de Paris a durablement institutionnalisé et administré les divergences plutôt qu’elle ne les a résolues. Le traitement national, le droit de priorité ou l’indépendance des brevets facilitent le brevetage multinational sans imposer l’harmonisation des conditions de délivrance et d’exploitation. Le caractère minimaliste de ce cadre constitue une condition de sa résilience : il permet aux États de coopérer tout en préservant leur autonomie. Le bilatéralisme et les arrangements restreints complètent ce dispositif.
La thèse analyse les acteurs de cette gouvernance, notamment par deux prosopographies sur les délégations aux conférences de révision et sur les membres de l’Association internationale pour la protection de la propriété industrielle. Autour du Bureau international et de cette association se constitue un milieu transnational de juristes, agents de brevets, fonctionnaires et industriels qui cherche à dépolitiser cet ordre par l’expertise. L’enquête, complétée par l’étude d’archives internationales et nationales, montre que malgré cette volonté de dépolitisation, les intérêts nationaux et le pouvoir des États restent décisifs.
Les deux guerres mondiales constituent enfin des moments privilégiés pour observer ces tensions. Elles révèlent la subordination possible des droits privés aux impératifs de guerre économique, de sécurité et de contrôle technologique, sans entraîner la disparition du cadre unioniste. La Seconde Guerre mondiale approfondit cette transformation en faisant de l’État un acteur direct de la production, de la gestion et de l’utilisation des inventions. Après 1945, la restauration de l’Union s’accompagne d’une fragmentation accrue de la gouvernance, d’une régionalisation européenne et d’un déplacement du centre de gravité vers les États-Unis. L’étude montre ainsi comment un ordre juridique faiblement contraignant a traversé guerres et mutations institutionnelles en maintenant un cadre commun sans abolir les souverainetés qui le limitaient.