Object Detection on STAR
14.5AP (Car)UHR-DETR
Evaluation results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| UHR-DETRBackbone=ResNet-50, Method Paradigm=Macro-level Patch Selection2026.04 | 14.5 | 10.4 | 34.6 | 33.9 | 52.4 | 20.2 | 70.1 | 43.2 | 34.9 | 15 | 36.4 | 40.5 | 0.357 | 3.13 | |
| CEASCBackbone=ResNet-50, Method Paradigm=Micro-level Dynamic Routing & SW2026.04 | 5.6 | 8.5 | 31.8 | 10.4 | 51.5 | 16.2 | 78.2 | 43.1 | 30.7 | 14.7 | 31.4 | 36.7 | 3.82 | 5 | |
| Faster-RCNNBackbone=ResNet-50, Method Paradigm=Exhaustive Sliding-Window Baselines (SW)2026.04 | 5.3 | 11.6 | 39 | 10.7 | 51 | 17.6 | 80.4 | 50.9 | 33.3 | 14.8 | 34.6 | 38.8 | 4.839 | 4.85 | |
| YOLOX-xBackbone=CSPDarknet, Method Paradigm=Exhaustive Sliding-Window Baselines (SW)2026.04 | 5.1 | 1.6 | 25.3 | 7.7 | 18.4 | 7.3 | 72.8 | 25.5 | 20.4 | 7.8 | 21.7 | 27.2 | 5.244 | 3.62 | |
| SparseFormerBackbone=SparseNet, Training=from scratch for 36 epochs, Method Paradigm=Micro-level Dynamic Routing & SW2026.04 | 4.9 | 0.6 | 25.2 | 5.8 | 44.8 | 12.7 | 74.2 | 17.3 | 23.2 | 7.6 | 23.8 | 30.4 | 4.146 | 2.51 | |
| RT-DETRBackbone=ResNet-50, Method Paradigm=Exhaustive Sliding-Window Baselines (SW)2026.04 | 4.7 | 10.9 | 30.9 | 7 | 62.3 | 16.2 | 80.5 | 44 | 32.1 | 14 | 33.1 | 37.7 | 3.068 | 2.32 | |
| FCOSBackbone=ResNet-50, Method Paradigm=Exhaustive Sliding-Window Baselines (SW)2026.04 | 4.6 | 6.5 | 35.4 | 7.3 | 50.2 | 14 | 79 | 38.6 | 29.5 | 12.5 | 31.1 | 36.6 | 3.384 | 2.27 | |
| RT-DETRBackbone=ResNet-50, Downsampling=1024x1024, Method Paradigm=Global Downsampling Approaches2026.04 | 1.4 | 1.8 | 15.8 | 8.9 | 6.8 | 14.3 | 39.2 | 21.8 | 13.8 | 0.6 | 9.5 | 31.4 | 0.036 | 0.46 | |
| SPDetBackbone=CSPDarknet, Method Paradigm=Macro-level Patch Selection2026.04 | 1.1 | 0.1 | 22.9 | 1.9 | 15.6 | 4.3 | 62.2 | 16.6 | 15.6 | 0.8 | 15.1 | 26.9 | 2.58 | 0.75 | |
| GigaDetBackbone=ResNet-50, Method Paradigm=Macro-level Patch Selection2026.04 | 1.1 | 1.2 | 20.8 | 3.7 | 23.5 | 1.3 | 50.7 | 11.5 | 14.2 | 4.6 | 15.2 | 17.8 | 0.345 | 3.12 | |
| FoveaBackbone=ResNet-50, Downsampling=1024x1024, Method Paradigm=Global Downsampling Approaches2026.04 | 0 | 0 | 12.6 | 4.3 | 3.3 | 8.2 | 35.2 | 19.3 | 10.3 | 0.2 | 6.4 | 24.8 | 1.56 | 1.97 |