Laser scanning technology allows acquisition of 3D point clouds of complex deformed geometries, with a millimeter precision. Steel structures, which are the primary focus of this paper, are susceptible to inelastic local buckling that, in turn, causes strength deterioration. Detailed point clouds afford an opportunity for automated assessment of the reserve capacity of steel beam-columns. With the overall goal of advancing the automated structural inspection methodologies, the objective of this paper is to test the performance of data driven models for prediction of reserve capacity of deformed steel members solely using point clouds as inputs. Specific focus is on developing a methodology that incorporates measurement noise present in the real-world point clouds into the approach for vision-based quantification of the reserve capacity. To this end, the STEEL‑3dPointClouds database of deformed steel column members obtained using physics-based simulations is leveraged for the development of the data-driven surrogates. Two machine learning models – namely, the k-nearest neighbors and a residual neural network – were rigorously evaluated. These methods solely rely on steel column profile point clouds. In the noise-free conditions, the residual neural network yielded promising results by achieving mean prediction errors of ∼2.9% and ∼1.5% for deformed shapes with reserve capacities above 40% and 80%, respectively. When measurement noise consistent with modern scanners (σ = 1 mm) is added to the point cloud, the prediction accuracy of the residual neural network remains comparable to the noise-free case. Moreover, even under a higher noise level of σ = 5 mm, the mean prediction errors of ∼3.8% and ∼2.1% are obtained for deformed shapes with reserve capacities above 40% and 80%, respectively. These findings demonstrate the potential of using the proposed methodology for automated inspection of steel structures.