Research map / 3D vision / Human pose estimation
CLOAF: CoLlisiOn-Aware Human Flow
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 CLOAF, a differentiable ODE-based flow in SMPL parameter space that removes body self-intersections from pose/shape estimates while keeping parametric body priors.
Problem
Feed-forward and optimization-based human pose and shape estimators often output bodies with self-intersections, and existing penalty-based fixes are non-differentiable and do not guarantee their absence.
Approach
Formulate body deformation between poses as an ODE in SMPL parameter space, with a neural velocity field projected onto the parametric space via the pseudo-inverse of the SMPL Jacobian; ODE diffeomorphism (Picard-Lindelöf) guarantees no self-intersections, and the whole pipeline is differentiable so it can be used for fine-tuning.
Key results
- CLOAF reduces the collision rate on 3DPW-test from 39.2% (HMR2.0) to 0.0% while keeping MPJPE at 82.3 mm versus 82.0 mm for HMR2.0.
- CLOAF reduces Accel.Err to 9.4 mm/s² from 16.1 for HMR2.0, versus 14.3-16.4 for optimization baselines.
- Fine-tuning EFT with differentiable CLOAF lowers P-MPJPE from 58.1 to 55.4 and collision rate from 6.5% to 3.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CLOAF collision correction to HMR2.0 predictions | HMR2.0 · 3DPW-test | Col.Rate@0: 39.2% -> 0.0% | Helped |
| Apply CLOAF collision correction to HMR2.0 predictions | HMR2.0 · 3DPW-test | Accel.Err: 16.1 -> 9.4 | Helped |
| Apply CLOAF collision correction to HMR2.0 predictions | HMR2.0 · 3DPW-test | P-MPJPE: 52.7 -> 53.2 | Hurt |
| Fine-tune EFT using differentiable CLOAF | EFT · 3DPW-test | P-MPJPE: 58.1 -> 55.4 | Helped |
| Fine-tune EFT using differentiable CLOAF | EFT · 3DPW-test | Col.Rate: 6.5 -> 3.4 | Helped |
| Use differentiable CLOAF rather than detaching its gradient (ablation) | EFT fine-tuning · 3DPW-test | P-MPJPE: 57.8 -> 55.4 | Helped |
| Use differentiable CLOAF rather than detaching its gradient (ablation) | EFT fine-tuning · 3DPW-test | Col.Rate: 6.3 -> 3.4 | Helped |
| Initialize CLOAF by randomly jittering intersecting body parameters rather than using successive frames (ablation) | CLOAF · 3DPW-test | MPJPE: 82.3 -> 81.6 | Helped |
| Initialize CLOAF from keyposes rather than using successive frames (ablation) | CLOAF · 3DPW-test | MPJPE: 82.3 -> 82.9 | Hurt |
| Initialize CLOAF from keyposes rather than using successive frames (ablation) | CLOAF · 3DPW-test | Accel.Err: 9.4 -> 16.4 | Hurt |
Surprising findings
- Fine-tuning with a detached (non-differentiable) CLOAF gives results comparable to the optimization-based method, showing the benefit comes from differentiability.
Limitations
- Some reconstruction accuracy is slightly reduced compared to baselines (P-MPJPE 53.2 vs 52.7 for HMR2.0).
- Requires a self-intersection-free starting pose; alternative initializations (Jitter, Keyposes) are needed for single images.
- Fine-tuning with CLOAF is less efficient at removing self-intersections than post-processing.
Experimental setup
Datasets: AMASS 3DPW COCO-EFT pseudo ground truth
Benchmarks: 3DPW
Metrics: MPJPE P-MPJPE Accel.Err Col.Rate
Baselines: HMR2.0 PARE TUCH EFT SPIN Opt.Cones (SMPLify-X) Opt.Contact (SMPLify-DC) COAP
Builds on: ShapeFlow NeuralODE SMPLify-X EFT AMASS
Closest papers by meaning
- PoseD-Flow: Versatile and Guided Flow Matching Model of Human Pose (CVPR 2026 · similarity 0.85)
- Trajectory Optimization for Physics-Based Reconstruction of 3D Human Pose From Monocular Video (CVPR 2022 · similarity 0.85)
- H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning (NeurIPS 2026 · similarity 0.85)
- Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.83)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024 · similarity 0.83)
- Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image (NeurIPS 2026 · similarity 0.83)
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.