Abstract

We propose a new single-shot method for multi-person 3D pose estimation ingeneral scenes from a monocular RGB camera. Our approach uses novelocclusion-robust pose-maps (ORPM) which enable full body pose inference evenunder strong partial occlusions by other people and objects in the scene. ORPMoutputs a fixed number of maps which encode the 3D joint locations of allpeople in the scene. Body part associations allow us to infer 3D pose for anarbitrary number of people without explicit bounding box prediction. To trainour approach we introduce MuCo-3DHP, the first large scale training data setshowing real images of sophisticated multi-person interactions and occlusions.We synthesize a large corpus of multi-person images by compositing images ofindividual people (with ground truth from mutli-view performance capture). Weevaluate our method on our new challenging 3D annotated multi-person test setMuPoTs-3D where we achieve state-of-the-art performance. To further stimulateresearch in multi-person 3D pose estimation, we will make our new datasets, andassociated code publicly available for research purposes.