ResiHMR Residual-limb aware 3D human mesh recovery
ResiHMR reconstructed meshes and representative limb-loss images

CVPR 2026 Highlight

ResiHMR: Residual-Limb Aware Single-Image 3D Human Mesh Recovery for Individuals with Limb Loss

Accepted to CVPR 2026 Highlight

Jiaying Ying*, Heming Du*, Kaihao Zhang, Sean M. Tweedy, Xin Yu

ResiHMR adapts single-image human mesh recovery to non-standard limb anatomy by optimizing over the observed kinematic topology and explicitly reconstructing residual-limb termination geometry.

Paper PDF GitHub Download LDPose-LimbLoss Dataset

Project video

Introduction to ResiHMR

Project introduction video for ResiHMR. If the embedded player is blocked in local preview, watch it on YouTube.

Abstract

ResiHMR: Residual-Limb Aware Single-Image 3D Human Mesh Recovery for Individuals with Limb Loss

Existing HMR pipelines inherit intact-limb body priors and often hallucinate missing distal segments. ResiHMR keeps the compactness of parametric mesh recovery while adapting it to observed residual-limb anatomy.

ResiHMR introduces a residual-limb aware framework for recovering anatomically coherent 3D human meshes from a single RGB image. It uses residual-limb keypoints to constrain the optimization to valid kinematic subgraphs and then reconstructs explicit limb-termination surfaces instead of collapsing or masking the missing anatomy.

  • Residual Anchor-Factor Optimization refines body pose, camera, shape, and residual-limb proportions while respecting limb-loss topology.
  • Residual-Limb Reconstruction removes distal geometry and generates a smooth, closed stump surface around the optimized residual endpoint.
  • The method is compatible with optimization and regression HMR backbones, including SMPLify-X and HSMR.
Synchronized qualitative comparison with recent HMR methods. All method renderings play at the same frame index for direct visual comparison.

Method

Overview of our ResiHMR Framework

ResiHMR is compatible with any HMR pipeline that outputs SMPL-X parameters (camera, pose, and shape), including both optimization-based and regression-based methods.

Overview of the ResiHMR framework
Given an input image, SMPL-X is initialized using intact 2D keypoints. Our Residual Anchor–Factor Optimization adapts the kinematic graph by refining anchor joints and residual-limb proportions under supervision of residual-limb 2D keypoints. The Residual-Limb Reconstruction module then removes distal limb geometry and generates a smooth, watertight stump surface, producing anatomically realistic residual-limb aware meshes
Overview of ResiHMR with HSMR initialization
Overview of our ResiHMR Framework with other existing HMR methods to initialize the SMPL-X/SMPL. In this case, we have HSMR as the example for the regression-based HMR method.
Algorithm 1 and Algorithm 2 for ResiHMR
Algorithm 1: Residual Anchor-Factor Optimization and Algorithm 2: Residual Limb Reconstruction.

Evaluation dataset

LDPose-LimbLoss Evaluation Dataset

The evaluation set is a curated subset of LDPose with added body masks for silhouette evaluation. It covers upper- and lower-limb loss, diverse activities, occlusions, prosthetic devices, and real-world backgrounds.

Each image includes 2D annotations of 17 standard body keypoints, 8 residual-limb endpoints, and per-person segmentation masks used to isolate the human body region for mesh evaluation.

Download Dataset LDPose Reference
Images and keypoints Refer back to the original LDPose release and source licensing.
Masks Released under CC BY-SA 4.0.
Representative samples from the LDPose-LimbLoss evaluation dataset
Representative LDPose-LimbLoss examples with green body-mDataset Demonstration of the proposed LDPose-LimbLoss Evaluation Dataset. Representative samples illustrating the diversity of subjects, amputation levels, poses, activities, and environments included in the dataset. Green overlays indicate the manually annotated body masks used to isolate the human body region.ask overlays.
Key statistics of the LDPose-LimbLoss evaluation dataset
Key Statistics of our LDPose-LimbLoss Evaluation Dataset. (a) Distribution of residual-limb types, covering upper- and lower-limb amputations across both sides of the body. (b) Gender distribution. (c) Ethnic composition of the participants. Together, these statistics demonstrate the dataset’s demographic and impairment diversity, providing a representative foundation for benchmarking residual-limb–aware 2D/3D human pose and mesh reconstruction.

QuantitativeResults

Comparison of recent HMR methods and ResiHMR

ResiHMR is evaluated with 2D reprojection error for body and residual-limb keypoints, plus mask IoU for silhouette consistency.

Method Body Kpts MPJPE ↓ Res-Limb MPJPE ↓ Intact Kpts MPJPE ↓ mIoU ↑
TokenHMR 34.79 102.34 31.73 0.717
CameraHMR 29.26 78.13 25.56 0.752
HSMR 28.27 73.61 24.56 0.705
SMPLify-X 47.67 129.59 41.32 0.662
ResiHMR (SMPLify-X) 41.77 98.36 37.40 0.703
ResiHMR (HSMR) 24.75 23.19 24.87 0.741

Numbers are reported on the LDPose-LimbLoss Evaluation Dataset. Lower MPJPE is better; higher mIoU is better. For all other HMR methods, as they do not explicitly predict residual-limb endpoints, we define a naive mid-point proxy on the corresponding limb segment to enable consistent scoring under a unified protocol.

Citation

Reference

Citation information can be updated after the final proceedings metadata is available. 

@misc{ying2026resihmr,
  title  = {ResiHMR: Residual-Limb Aware Single-Image 3D Human Mesh Recovery for Individuals with Limb Loss},
  author = {Ying, Jiaying and Du, Heming and Zhang, Kaihao and Tweedy, Sean M. and Yu, Xin},
  year   = {2026}
}