Research map / 3D vision / Human pose estimation
Learning Skeletal Graph Neural Networks for Hard 3D Pose Estimation
Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, Qiang Xu
Read the paperCheck your idea against 137,217 papers
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 hierarchical fusion of short- and long-range hops instead of single-stage fusion (ablation) | HCSF skeletal GNN · Human3.6M (2D GT) | MPJPE (mm): outperforms Non-hierarchy by 2mm~3.7mm under different L | Improved |
| add hop-aware (per-hop channel squeezing) processing to the hierarchy (ablation) | HCSF skeletal GNN · Human3.6M (2D GT) | MPJPE (mm): 34.2 -> 32.6 (L=3); consistent drop of 1.6mm~1.8mm without it | Improved |
| reduce squeezing ratio d from 1 to 1/8 (ablation) | HCSF · Human3.6M (2D GT) | MPJPE (mm): 35.4 -> 31.4 | Improved |
| squeeze further to d=1/16 (ablation) | HCSF · Human3.6M (2D GT) | MPJPE (mm): 31.4 -> 32.6 | Hurt |
| initialise learnable graph from physical topology rather than dense or random (ablation) | dynamic graph learning · Human3.6M (2D GT) | MPJPE (mm): physical 32.1 vs dense 35.7 vs random 40.8 | Improved |
| learn graph only from dynamic offsets of input features (ablation) | dynamic graph learning · Human3.6M (2D GT) | MPJPE (mm): 30.8 -> 44.3 | Hurt |
| add dynamic offsets to weighted graph M_k (Eq. 8 with alpha) (ablation) | HCSF static baseline · Human3.6M (2D GT) | MPJPE (mm): 30.8 -> 30.5 (M+O), 30.4 (Eq. 8) | Improved |
| add temporal-aware scheme to dynamic graph construction (ablation) | HCSF with dynamic graph · Human3.6M (2D GT) | MPJPE (mm): 30.8 -> 29.7 | Improved |
| add temporal-aware scheme to dynamic graph (detected 2D temporal input) (ablation) | Ours · Human3.6M (temporal, detected) | MPJPE (mm): 46.4 -> 45.7 | Improved |
| use 3x1 temporal kernel instead of longer or strided kernels (ablation) | temporal-aware dynamic graph · Human3.6M (2D GT) | MPJPE (mm): 30.8 -> 30.4 (3,1); 30.7 (7,1) | Mixed |
| replace k-hop, MixHop or MS-Hop aggregation with HCSF (ablation) | skeleton action recognition GNN · NTU RGB+D X-View (static graph) | top-1 accuracy (%): 94.1 -> 94.8 (vs MS-Hop) | Improved |
| use 3-frame temporal convolution in dynamic graph for action recognition (ablation) | HCSF dynamic graph · NTU RGB+D X-View | top-1 accuracy (%): 94.7 -> 95.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SRNet — SRNet: Improving Generalization in 3D Human Pose Estimation with a Split-and-Recombine Approach (ECCV 2020) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Graph Convolutional Network for Skeleton-Based Human Action Recognition by Neural Searching (AAAI 2020 · similarity 0.90)
- Topology-Aware Convolutional Neural Network for Efficient Skeleton-Based Action Recognition (AAAI 2022 · similarity 0.89)
- Modulated Graph Convolutional Network for 3D Human Pose Estimation (ICCV 2021 · similarity 0.89)
- Context Aware Graph Convolution for Skeleton-Based Action Recognition (CVPR 2020 · similarity 0.89)
- BlockGCN: Redefine Topology Awareness for Skeleton-Based Action Recognition (CVPR 2024 · similarity 0.88)
- Dynamic Semantic-Based Spatial Graph Convolution Network for Skeleton-Based Human Action Recognition (AAAI 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.
