D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement

Benchmark Model Rank Results
real-time-object-detection-on-coco-val2017D-FINE-X+ (Objects365 pre-training)#4box AP: 59.3FPS (batch 1, GPU in brackets): 78 (T4)
real-time-object-detection-on-coco-val2017D-FINE-L+ (Objects365 pre-training)#10box AP: 57.1FPS (batch 1, GPU in brackets): 124 (T4)
real-time-object-detection-on-coco-val2017D-FINE-X#16box AP: 55.8FPS (batch 1, GPU in brackets): 78 (T4)
real-time-object-detection-on-coco-val2017D-FINE-M+ (Objects365 pre-training)#19box AP: 55.1FPS (batch 1, GPU in brackets): 178 (T4)
real-time-object-detection-on-coco-val2017D-FINE-L#30box AP: 54.0FPS (batch 1, GPU in brackets): 124 (T4)
real-time-object-detection-on-coco-val2017D-FINE-M#60box AP: 52.3FPS (batch 1, GPU in brackets): 180 (T4)
real-time-object-detection-on-coco-val2017D-FINE-S+ (Objects365 pre-training)#77box AP: 50.7FPS (batch 1, GPU in brackets): 287 (T4)
real-time-object-detection-on-coco-val2017D-FINE-S#93box AP: 48.5FPS (batch 1, GPU in brackets): 287 (T4)