Research map / 3D vision / Human pose estimation
KITRO: Refining Human Mesh by 2D Clues and Kinematic-tree Rotation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KITRO, a closed-form, kinematic-tree decision-tree refinement that uses 2D keypoints, bone lengths and parent depth to refine SMPL human meshes from HMR base models.
Problem
Refining estimated 3D human meshes with 2D keypoints via reprojection-loss gradient descent suffers from depth ambiguity and divergent gradients from distal joints that harm proximal joints along the kinematic chain.
Approach
Computes two closed-form child-bone direction hypotheses (toward/away from camera) from 2D keypoints, bone length and parent depth, organizes them into a binary decision tree along the kinematic tree, and selects the most probable path by softmax cosine similarity with the initial HMR bone rotations; shape is refined via a 2D bone-length loss and camera via least squares with moving average.
Key results
- On 3DPW with CLIFFb base, KITRO reaches 27.67 PA-MPJPE, 43.53 MPJPE and 53.44 PVE, versus 32.04/55.83/71.95 for CLIFFr refinement.
- On Human3.6M, KITRO achieves 21.04 PA-MPJPE and 34.50 MPJPE, versus 25.88/42.79 for CLIFFr.
- On FreiHAND with MANO, KITRO reduces PA-MPJPE from 7.82 (baseline) to 4.06, compared to 4.66 for SMPLify.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply KITRO refinement to CLIFFb | CLIFFb · 3DPW | PA-MPJPE: 32.04 -> 27.67 | Helped |
| apply KITRO refinement to CLIFFb | CLIFFb · 3DPW | MPJPE: 55.83 -> 43.53 | Helped |
| apply KITRO refinement to CLIFFb | CLIFFb · Human3.6M | PA-MPJPE: 25.88 -> 21.04 | Helped |
| apply KITRO refinement to CLIFFb | CLIFFb · Human3.6M | MPJPE: 42.79 -> 34.50 | Helped |
| add camera translation refinement (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 28.53 -> 27.67 | Helped |
| use moving-average camera translation updates (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 27.99 -> 27.67 | Helped |
| use the full-perspective focal length instead of a fixed focal length of 5000 (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 32.72 -> 27.67 | Helped |
| refine shape parameters (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 35.00 -> 27.67 | Helped |
| optimize shape using projected bone-length loss instead of 2D reprojection loss (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 32.60 -> 27.67 | Helped |
| refine pose parameters (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 44.57 -> 27.67 | Helped |
| use decision-tree hypothesis selection instead of greedy selection (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 29.24 -> 27.67 | Helped |
| use weighted soft pose updates instead of hard updates (ablation) | KITRO on CLIFFb · 3DPW | PA-MPJPE: 32.18 -> 27.67 | Helped |
Surprising findings
- Competing refinement methods (SMPLify, CLIFFr) can degrade proximal-joint improvement along the arm kinematic chain, while KITRO improves all joints.
- Removing the pose refinement increases errors by roughly 60%, larger than the effect of shape (~30%) or camera (<1%).
Limitations
- Relies on the initial HMR mesh for hypothesis selection, so poor initial predictions (e.g., person misidentification) lead to deviations.
- Depends on the accuracy of input 2D keypoints.
- Evaluated using ground-truth 2D keypoints for main results.
Experimental setup
Datasets: 3DPW Human3.6M FreiHAND
Benchmarks: 3DPW Human3.6M FreiHAND
Metrics: MPJPE PA-MPJPE PVE
Baselines: SMPLify CLIFFr DynaBOA Pose2Mesh SPIN + SMPLify EFTb + SMPLify PRoM + SMPLify
Builds on: SMPL SMPLify CLIFF SPIN EFT
Closest papers by meaning
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · similarity 0.84)
- REACTO: Reconstructing Articulated Objects from a Single Video (CVPR 2024 · similarity 0.84)
- SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons (CVPR 2025 · similarity 0.83)
- FlexPose: Pose Distribution Adaptation with Limited Guidance (AAAI 2025 · similarity 0.83)
- Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization (NeurIPS 2025 · similarity 0.82)
- PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.