Light3DPose: Real-time Multi-Person 3D PoseEstimation from Multiple Views

We present an approach to perform 3D pose estimation of multiple people from\na few calibrated camera views. Our architecture, leveraging the recently\nproposed unprojection layer, aggregates feature-maps from a 2D pose estimator\nbackbone into a comprehensive representation of the 3D scene. Such intermediate\nrepresentation is then elaborated by a fully-convolutional volumetric network\nand a decoding stage to extract 3D skeletons with sub-voxel accuracy. Our\nmethod achieves state of the art MPJPE on the CMU Panoptic dataset using a few\nunseen views and obtains competitive results even with a single input view. We\nalso assess the transfer learning capabilities of the model by testing it\nagainst the publicly available Shelf dataset obtaining good performance\nmetrics. The proposed method is inherently efficient: as a pure bottom-up\napproach, it is computationally independent of the number of people in the\nscene. Furthermore, even though the computational burden of the 2D part scales\nlinearly with the number of input views, the overall architecture is able to\nexploit a very lightweight 2D backbone which is orders of magnitude faster than\nthe volumetric counterpart, resulting in fast inference time. The system can\nrun at 6 FPS, processing up to 10 camera views on a single 1080Ti GPU.\n