GAT vs Pure Hierarchical Architecture Comparison
GAT: Graph Attention with learnable edge weights
GCN: Fixed graph convolution
GAT: Multi-path with global fusion
GCN: Single-path sequential
GAT: 27% better MPJPE
GCN: Faster inference
Model | Architecture | MPJPE (mm) | P-MPJPE (mm) | OKS mAP (%) | OKS@0.5 | OKS@0.9 | Std Dev |
---|---|---|---|---|---|---|---|
HierGAT_7h8_g2 | Hier+GAT | 0.2595 | 0.2106 | 99.75% | 100.00% | 99.65% | 0.3058 |
Hier_7h16 | Pure Hier | 0.2908 | 0.2424 | 99.71% | 99.94% | 99.36% | 0.3014 |
HierGAT_7h4_g8 | Hier+GAT | 0.2922 | 0.2370 | 99.64% | 99.94% | 99.36% | 0.3041 |
Hier_8h8 | Pure Hier | 0.3272 | 0.2542 | 99.60% | 99.94% | 99.18% | 0.3152 |
Hier_7h4 | Pure Hier | 0.4302 | 0.3913 | 99.50% | 100.00% | 98.84% | 0.2977 |
Hier_8h4 | Pure Hier | 0.4780 | 0.3964 | 99.41% | 99.94% | 98.66% | 0.3690 |