Hybrid-RT-DETR

Hybrid encoder-decoder network for end-to-end object detection in UAV imagery

基于多尺度增强辅助编解码网络的无人机目标检测算法
Keywords：DETR；多尺度特征融合；辅助目标查询

Results on VisDrone2019

Method	mAP	$AP_{50}$	$AP_{75}$	$AP_s$	$AP_m$	$AP_l$
YOLOv5m6	27.00	44.4	—	—	—	—
YOLOv6m	20.90	35.50	20.81	10.43	33.21	57.64
YOLOv7	28.74	49.91	—	—	—	—
YOLOv8m	27.33	44.71	—	—	—	—
DAB-deformable-detr-R50	25.72	43.51	25.73	17.30	36.74	45.82
DN-DAB-deformable-detr-R50	27.93	46.00	28.41	19.14	39.67	54.31
Lite-DINO-H3L1-R50	32.79	55.21	32.70	23.38	44.65	58.81
RT-DETR-R18	28.93	48.29	29.21	19.55	40.43	61.00
RT-DETR-R50	32.20	52.85	32.65	22.27	44.87	68.43
本章算法-R18	30.57	50.78	30.91	21.40	42.38	59.70
本章算法-R50	33.14	54.43	33.89	23.62	45.94	63.08

模型在TITAN RTX上的推理耗时 FP32/wo TensorRT

模型	平均推理耗时（秒）
baseline_r18	0.027
ours_r18	0.048
baseline_r50	0.048
ours_r50	0.055

---

Quick start

Install

pip install -r requirements.txt

Data

Download VisDrone and convert it to COCO format annonations of train and val data.

Training & Evaluation

• Training on a Single GPU:

# training on single-gpu
export CUDA_VISIBLE_DEVICES=0
python tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml

• Training on Multiple GPUs:

# train on multi-gpu
export CUDA_VISIBLE_DEVICES=0,1,2,3
torchrun --nproc_per_node=4 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml

• Evaluation on Multiple GPUs:

# val on multi-gpu
export CUDA_VISIBLE_DEVICES=0,1,2,3
torchrun --nproc_per_node=4 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml -r path/to/checkpoint --test-only

Test

python tools/infer.py

Tips: set remap_mscoco_category: False.