Research map / 3D vision / Human pose estimation
$\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation
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 Di²Pose, a two-stage framework that quantizes 3D human poses into discrete tokens (FSQ with Local-MLP encoder) and models them with a conditional discrete diffusion process using occlude/replace transitions for occluded monocular 3D pose estimation.
Problem
Continuous diffusion models for monocular 3D human pose estimation need large search spaces and data, producing biomechanically implausible poses, especially under occlusion.
Approach
Represent 3D poses as compositional discrete tokens via a VQ-VAE-like pose quantization step with Local-MLP blocks and FSQ, then run a discrete diffusion over tokens with an occlude-and-replace transition matrix conditioned on image features, so occlusion is modeled in latent space.
Key results
- Achieves 49.2mm average MPJPE on Human3.6M, 0.5mm better than the prior best (Gong et al., 49.7mm).
- On 3DPW, reaches 79.3mm MPJPE / 50.1mm PA-MPJPE, beating Gong et al. by 3.4mm MPJPE and 3.7mm PA-MPJPE.
- On 3DPW-AdvOcc@80, MPJPE 153.6mm vs 189.3mm for the next best method (Gong et al.).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the strongest reported baseline with Di²Pose | 3D pose estimation · Human3.6M | MPJPE: 49.7 -> 49.2 | Helped |
| Replace the strongest reported baseline with Di²Pose | 3D pose estimation · Human3.6M | PA-MPJPE: 39.2 -> 39.0 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW | MPJPE: 82.7 -> 79.3 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW | PA-MPJPE: 53.8 -> 50.1 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW-Occ | MPJPE: 82.1 -> 79.6 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW-Occ | PA-MPJPE: 53.5 -> 50.7 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW-AdvOcc@40 | MPJPE: 121.4 -> 108.4 | Helped |
| Replace the strongest baseline with Di²Pose | 3D pose estimation · 3DPW-AdvOcc@80 | MPJPE: 189.3 -> 153.6 | Helped |
| Add the Local-MLP pose representation instead of PCT (ablation) | Pose quantization step · Human3.6M pose reconstruction | MPJPE: 15.2 -> 13.6 | Helped |
| Use Joint Shift with X=3 rather than X=1 (ablation) | Pose quantization step · Human3.6M pose reconstruction | MPJPE: 14.5 -> 13.6 | Helped |
| Use Joint Shift with X=3 rather than X=5 (ablation) | Pose quantization step · Human3.6M pose reconstruction | MPJPE: 14.1 -> 13.6 | Helped |
| Use FSQ levels [7,5,5,5,5] rather than [8,5,5,5] (ablation) | Pose quantization step · Human3.6M pose reconstruction | MPJPE: 15.2 -> 13.6 | Helped |
Surprising findings
- Using fewer inference steps (25 instead of 100) at test time only modestly degrades MPJPE (50.3 vs 49.2 mm with 100 training steps).
Limitations
- Predictions can deviate from ground truth under severe occlusions that confound human observers.
- Designed only for frame-based estimation; does not use inter-frame video information.
- Monocular 2D-to-3D inference lacks depth information, causing inherent uncertainty.
Experimental setup
Datasets: Human3.6M 3DPW 3DPW-Occ DTD
Benchmarks: Human3.6M 3DPW 3DPW-Occ 3DPW-AdvOcc
Metrics: MPJPE PA-MPJPE
Baselines: DiffPose Gong et al. (CVPR'23) Geng et al. (CVPR'23) Zhang et al. (TPAMI'23) Choi et al. (IROS'23) Cai et al. (ICCV'19) Pavllo et al. (CVPR'19) Zheng et al. (ICCV'21)
Builds on: VQ-VAE finite scalar quantization discrete diffusion (D3PM-style) DiffPose MLP-Mixer
Closest papers by meaning
- Dual-Diffusion for Binocular 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.91)
- Disentangled Diffusion-Based 3D Human Pose Estimation with Hierarchical Spatial and Temporal Denoiser (AAAI 2024 · similarity 0.89)
- HiPART: Hierarchical Pose AutoRegressive Transformer for Occluded 3D Human Pose Estimation (CVPR 2025 · similarity 0.88)
- Object Pose Estimation via the Aggregation of Diffusion Features (CVPR 2024 · similarity 0.88)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.88)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.87)
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.