RT-YODETR: Modified RT-DETR with YOLO (v9) backbone

Official RT-DETR repository: https://github.com/lyuwenyu/RT-DETR

Official YOLOv9 repository: https://github.com/WongKinYiu/yolov9

PAPER SOON PUBLISHED

---

Figures

	Model	mAP 50:95	Params	Inference time on 1 image RTX-3050 Mobile	Epoch
2	RT-DETR-L-YOLOv9Ebb	0.502	48400298	0.1297	27
3	RT-DETR-L-R50	0.42	42702570	0.1124	27
4	RT-DETR-L-R101	0.4333	76395978	0.1516	42

The existing weight is trained with trimmed COCO dataset This research is quite constrained with the budget, so to train enough models, we have to trim the COCO dataset with specific below:

Stage	Class/label	Total image count	Total annotation count	Image count used in this training	Annotation count used in this training
Training	person	64115	269578	5000	20753
Training	bicycle	3252	269578	3252	23261
Validation	person	2693	11320	2693	11279
Validation	bicycle	149	11320	149	1118

Modification & Training Procedure

• Removed auxiliary branch from YOLOv9 E Backbone
• Plug in the backbone into the RT-DETR framework
• Load the YOLOv9 E backbone weight into the new RT-DETR backbone
• Freeze the backbone
• Train & Benchmarking the model

Citation

Paper publication ASAP