TinyFormer: Preserving Tiny Objects in YOLO-DETR Hybrid Real-time Detectors

Benchmark Model Rank Results
real-time-object-detection-on-coco-val2017TinyFormer-XL-PBM (Objects365 pre-training)#1box AP: 62.5FPS (batch 1, GPU in brackets): 126 (RTX 3090)
real-time-object-detection-on-coco-val2017TinyFormer-XL-PBM#2box AP: 60.6FPS (batch 1, GPU in brackets): 126 (RTX 3090)
real-time-object-detection-on-coco-val2017TinyFormer-X#5box AP: 58.4FPS (batch 1, GPU in brackets): 216 (RTX 3090)
real-time-object-detection-on-coco-val2017TinyFormer-L-PBM#11box AP: 56.8
real-time-object-detection-on-coco-val2017TinyFormer-L#14box AP: 56.5FPS (batch 1, GPU in brackets): 282 (RTX 3090)
real-time-object-detection-on-coco-val2017TinyFormer-M-PBM#39box AP: 53.7
real-time-object-detection-on-coco-val2017TinyFormer-M#40box AP: 53.5FPS (batch 1, GPU in brackets): 328 (RTX 3090)
real-time-object-detection-on-coco-val2017TinyFormer-S-PBM#73box AP: 51.5
real-time-object-detection-on-coco-val2017TinyFormer-S#75box AP: 51.3FPS (batch 1, GPU in brackets): 429 (RTX 3090)
object-detection-on-coco-val2017TinyFormer-XL-PBM#24box AP: 60.6
object-detection-on-coco-val2017TinyFormer-X-PBM#51box AP: 58.5
object-detection-on-coco-val2017TinyFormer-L-PBM#78box AP: 56.8
object-detection-on-coco-val2017TinyFormer-M-PBM#137box AP: 53.7
object-detection-on-coco-val2017TinyFormer-S-PBM#187box AP: 51.5