Scanpath Prediction on COCO-Search18 (test)
50.4Scanpath Accuracy (R@1)Gazeformer
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GazeformerScanpath length (fixations)=62026.04 | 50.4 | 49 | 1.928 | 2.072 | — | — | |
| CLIPgazeScanpath length (fixations)=62026.04 | 47.6 | 54.5 | 1.489 | 2.014 | — | — | |
| HATScanpath length (fixations)=62026.04 | 46.8 | 54 | 1.522 | 2.063 | — | — | |
| Human ConsistencyScanpath length (fixations)=62026.04 | 46.3 | 47 | 2.144 | 2.353 | — | — | |
| SemBA × DETRFovea Configuration=Multi-Scale Fovea (4 × 160 pixel), Scanpath length (fixations)=62026.04 | 42.6 | 48.8 | 2.102 | 2.574 | — | — | |
| SemBA × RT-DETRFovea Configuration=Multi-Scale Fovea (4 × 160 pixel), Scanpath length (fixations)=62026.04 | 42.1 | 47.3 | 2.167 | 2.635 | — | — | |
| SemBA × YOLOv11Fovea Configuration=Multi-Scale Fovea (4 × 160 pixel), Scanpath length (fixations)=62026.04 | 41.4 | 44.8 | 2.161 | 2.572 | — | — | |
| ISP-SENetadaptation=embedding-based2025.04 | 35.25 | — | — | — | — | — | |
| IVSNScanpath length (fixations)=62026.04 | 34.6 | 38 | 3.034 | 3.36 | — | — | |
| ChenLSTM-ISPadaptation=standard2025.04 | 34.28 | — | — | — | — | — | |
| Gazeformer-ISP-Sadaptation=subject-specific2025.04 | 33.53 | — | — | — | — | — | |
| ChenLSTM-ISP-Sadaptation=subject-specific2025.04 | 33.17 | — | — | — | — | — | |
| Gazeformer-ISPadaptation=standard2025.04 | 31.73 | — | — | — | — | — | |
| Clean ModelTrigger=None, Poisoning ratio (ρ)=0%2026.04 | — | — | — | — | 86.6 | 86.4 | |
| Fixed-Path Backdoor AttackTrigger=Vision, Poisoning ratio (ρ)=10%2026.04 | — | — | — | — | 81.4 | 0.8 | |
| Fixed-Path Backdoor AttackTrigger=Vision, Poisoning ratio (ρ)=5%2026.04 | — | — | — | — | 82.4 | 1.5 | |
| Fixed-Path Backdoor AttackTrigger=Vision, Poisoning ratio (ρ)=2.5%2026.04 | — | — | — | — | 79.2 | 5.7 | |
| Fixed-Path Backdoor AttackTrigger=Language, Poisoning ratio (ρ)=10%2026.04 | — | — | — | — | 80.7 | 0 | |
| Fixed-Path Backdoor AttackTrigger=Language, Poisoning ratio (ρ)=5%2026.04 | — | — | — | — | 81.9 | 0 | |
| Fixed-Path Backdoor AttackTrigger=Language, Poisoning ratio (ρ)=2.5%2026.04 | — | — | — | — | 79.2 | 0.8 | |
| Fixed-Path Backdoor AttackTrigger=Vision-Language, Poisoning ratio (ρ)=10%2026.04 | — | — | — | — | 79.7 | 0.2 | |
| Fixed-Path Backdoor AttackTrigger=Vision-Language, Poisoning ratio (ρ)=5%2026.04 | — | — | — | — | 82.4 | 0.3 | |
| Fixed-Path Backdoor AttackTrigger=Vision-Language, Poisoning ratio (ρ)=2.5%2026.04 | — | — | — | — | 81.5 | 0.7 |