Abstract
Monocular 3D human mesh recovery (HMR) is an essential perception capability for human-centered embodied AI and clinical movement analysis. Most top-down HMR pipelines normalize a detected person crop to recover full-image camera geometry. However, under unconstrained occlusion or image truncation, detectors often localize only the visible body region. This fundamentally biases the crop-to-image camera conversion toward a visible-body frame rather than the desired full-body frame, propagating severe spatial distortions downstream. To address this, we propose FRAME (Full-body Recovery via Adaptive Margin Extension), an upstream geometric rectification framework. Conditioned on keypoint confidence patterns and image-boundary cues, AME predicts direction-specific expansion ratios to proactively recover an asymmetric full-body localization frame before mesh regression. The recovered bounding box, allowed to extend beyond image boundaries under truncation, provides geometrically consistent inputs for downstream HMR. On 3DOH50K, FRAME reduces the Mean Per-Joint Position Error (MPJPE) and Mean Per-Vertex Error (MPVE) by 10.5 mm and 9.7 mm over the baseline, respectively, while showing consistent gains in controlled component ablations. These results support full-body frame recovery as an effective geometric strategy for robust top-down HMR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CREATE_001.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=7QBdcnqA8U
BibTex
@InProceedings{ZhaNan_FRAME_MICCAISAT2026,
author = { Zhang, Nan AND Jin, Zirong AND Zhang, Shuo AND Liu, Hongbin},
title = { { FRAME: Full-body Recovery via Adaptive Margin Extension for Occlusion-Robust 3D Human Pose and Shape Estimation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}