MVMP-HMR: Multiview Multi-Person Human Mesh Recovery Under Large Scenes with Occlusions

ICLR 2026 Conference Submission12958 Authors

Keywords: human mesh recovery, multiview, 3D pose

Abstract: Human mesh recovery (HMR) refers to recovering the human 3D meshes from images. Most existing HMR tasks focus on multi-person from a single image or a single person from multiple views. And the evaluation benchmarks used in these methods usually contain quite small numbers of humans or under small scenes, which is unreliable for real applications with severe occlusions. Thus, we present Multiview Multi-Person HMR (MVMP-HMR), a multiview model for multi-person whole-body human mesh recovery from multi-view images under occluded scenes. Specifically, MVMP-HMR first fuses multiple views to obtain a 3D feature volume for all persons, and then the pelvis joint from a 3D pose estimation net is utilized to acquire the human query of each person from the 3D feature volume. Finally, the human queries are cross-attentioned with the 3D feature volume and integrated to decode each person's 3D meshes. Besides, two novel losses are put forward to further enhance the model performance: the orientation loss and the 3D joint density loss, dealing with the orientation and pose ambiguities in the mesh predictions under the occluded scenes. Furthermore, a large synthetic MVMP-HMR dataset is proposed, which consists of 15 multiview scenes with up to 50 camera views and 30 persons. Experiments demonstrate that the existing state-of-the-art (SOTA) HMR methods cannot perform well on the proposed large MVMP-HMR benchmark, and the proposed MVMP-HMR model's advantages over existing SOTAs under large scenes with severe occlusions.

Primary Area: datasets and benchmarks

Submission Number: 12958