Research map / 3D vision / Human pose estimation
Unsupervised Geometry-Aware Representation for 3D Human Pose Estimation
Helge Rhodin, Mathieu Salzmann, Pascal Fua
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometry-aware 3D latent space (ablation) | UNet encoder-decoder · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 191.7 -> 145.6 | Improved |
| add geometry-aware 3D latent space (ablation) | UNet encoder-decoder · Human3.6M S1 (UNet, no rotation aug.) | P-MPJPE (mm): 139.0 -> 112.2 | Improved |
| add separate appearance latent space (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 159.0 -> 145.6 | Improved |
| add separate appearance latent space (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | P-MPJPE (mm): 117.1 -> 112.2 | Improved |
| add separate background handling (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 159.6 -> 145.6 | Improved |
| add separate background handling (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | P-MPJPE (mm): 124.6 -> 112.2 | Improved |
| use deconvolution instead of bilinear upsampling (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 149.2 -> 145.6 | Improved |
| add ResNet-feature (ImageNet) loss (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 154.1 -> 145.6 | Improved |
| use two hidden layers in pose decoder F instead of one (ablation) | UNet · Human3.6M S1 (UNet, no rotation aug.) | N-MPJPE (mm): 157.4 -> 145.6 | Improved |
| use ResNet encoder instead of UNet vs the baseline semi-supervised method | ResNet-50 encoder · Human3.6M S1 | MPJPE: 30.7mm (20 % relative improvement) | Improved |
| use unsupervised geometry-aware representation vs baseline semi-supervised method | UNet · Human3.6M S1 | MPJPE: 17.4mm (11.3% relative improvement) | Improved |
| use fewer than 180,000 labeled images | Ours vs baselines · Human3.6M | MPJPE | Mixed |
Closest papers by meaning
- Weakly-Supervised Discovery of Geometry-Aware Representation for 3D Human Pose Estimation (CVPR 2019 · similarity 0.93)
- Geometry-Driven Self-Supervised Method for 3D Human Pose Estimation (AAAI 2020 · similarity 0.90)
- Kinematic-Structure-Preserved Representation for Unsupervised 3D Human Pose Estimation (AAAI 2020 · similarity 0.89)
- Weakly-Supervised 3D Human Pose Learning via Multi-View Images in the Wild (CVPR 2020 · similarity 0.89)
- Harvesting Multiple Views for Marker-Less 3D Human Pose Annotations (CVPR 2017 · similarity 0.89)
- Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (NeurIPS 2024 · similarity 0.88)
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.
