Table 6 Ablation study on threshold granularity and learning strategy (RegDB V2I). Fixed baselines use manually set values, while learnable approaches optimize thresholds during training.
From: Dynamic adaptive synergistic attention network for visible-infrared person re-identification
Threshold Strategy | Params | Rank-1 (%) | mAP (%) | Limitation |
|---|---|---|---|---|
Global \(\tau = \text {mean}({\hat{W}})\) | 0 | 94.85 | 89.23 | Scale mismatch |
Uniform per-layer (fixed \(\tau _l\!=\!0.4\)) | 0 | 95.23 | 90.45 | No adaptation |
Per-layer \(\tau _l\) (learned, ours) | 4 | 96.20 | 92.12 | None (optimal) |
Per-channel \(\tau _c\) (learned) | 8192 | 95.14 | 90.67 | Overfitting |
