RF-DETR: Neural Architecture Search for Real-Time Detection Transformers

Benchmark Model Rank Results
object-detection-on-coco-val2017RF-DETR (Max)#19box AP: 61.8
object-detection-on-coco-val2017RF-DETR (2XL) (fine-tuned)#30box AP: 60.2
object-detection-on-coco-val2017RF-DETR (XL) (fine-tuned)#42box AP: 58.9
object-detection-on-coco-val2017RF-DETR (L) (fine-tuned)#84box AP: 56.5
object-detection-on-coco-val2017RF-DETR (Medium) (fine-tuned)#115box AP: 54.7
object-detection-on-coco-val2017RF-DETR (Small) (fine-tuned)#157box AP: 53.0
object-detection-on-coco-val2017RF-DETR (Nano) (fine-tuned)#257box AP: 48.4
real-time-object-detection-on-coco-val2017RF-DETR-2XL (880px) (Objects365 pre-training)–box AP: 60.1FPS (batch 1, GPU in brackets): 58 (T4)
real-time-object-detection-on-coco-val2017RF-DETR-L (704px) (Objects365 pre-training)–box AP: 56.5FPS (batch 1, GPU in brackets): 147 (T4)
real-time-object-detection-on-coco-val2017RF-DETR-M (576px) (Objects365 pre-training)–box AP: 54.7FPS (batch 1, GPU in brackets): 227 (T4)
real-time-object-detection-on-coco-val2017RF-DETR-N (384px) (Objects365 pre-training)–box AP: 48.0FPS (batch 1, GPU in brackets): 435 (T4)
real-time-object-detection-on-coco-val2017RF-DETR-S (512px) (Objects365 pre-training)–box AP: 52.9FPS (batch 1, GPU in brackets): 286 (T4)
real-time-object-detection-on-coco-val2017RF-DETR-XL (700px) (Objects365 pre-training)–box AP: 58.6FPS (batch 1, GPU in brackets): 87 (T4)