Supplementary material:

Direct Prediction of 3D Body Poses from Motion Compensated Sequences

CVPR 2016 Submission

Paper ID 166

We provide additional analysis and experimental results in the following supplementary text.

We can disambiguate challenging poses with mirroring and self-oclusion and achieve state-of-the-art performance by combining appearance and motion cues from rectified spatio-temporal volumes (RSTVs). We provide several example videos below. Best viewed in full-screen mode.

Buying

Eating

Walking Dog

The proposed framework is robust to ambiguities caused by the projection of 3D data to 2D.

Our method can disambiguate poses with significant amount of self-occlusion.

Using appearance and motion informaiton simultaneously our approach can reliably handle mirror ambiguities.

We obtain RSTVs using our CNN-based motion compensation algorithm. The video below depicts several motion compensation examples on our datasets.

We provide examples of 3D human pose estimation with KRR, KDE and DN regressors, applied on rectified spatio-temporal volumes. RSTV+DN yields more accurate 3D pose estimates.

We provide further visualization for the HumanEva dataset below.

The supplementary videos are encoded by FFMPEG with h.264 codec. If you can't play the video, please download the VLC player at: http://www.videolan.org/vlc/index.html