FLDet: A Faster and Lighter Aerial Object Detector

Benchmark

VisDrone2019

Model	Size	AP(%)	AP50(%)	#Params.	FLOPs	Latency	FPS	Weight&train.log
FLDet-N	640	16.7	30.1	1.2M	12.3G	17.9ms	55.9	Google Drive
FLDet-S	640	18.8	33.4	2.4M	26.9G	24.3ms	41.2	Google Drive

UAVDT

Model	Size	AP(%)	AP50(%)	#Params.	FLOPs	Latency	FPS	Weight&train.log
FLDet-N	640	16.8	28.8	1.2M	12.3G	17.8ms	56.2	Google Drive
FLDet-S	640	17.5	30.3	2.4M	26.9G	24.2ms	41.3	Google Drive

Code

The repo is the official implementation of FLDet.

Our config file is at ultralytics/cfg/models/FLDet

Requirement

1. Install torch 2.0.1 and torchvision 0.15.2

pip install torch==2.0.1 torchvision==0.15.2 --index-url https://download.pytorch.org/whl/cu118

2. Install other requirements

pip install -e .

Usage

Data preparation

You could download dataset form VisDrone(YOLO Format) and UAVDT dataset (YOLO Format) .

Training

1. VisDrone2019

% FLDet-N
yolo detect train data=VisDrone_test.yaml model=FLDet-N.yaml imgsz=640 device=0,1,2,3 optimizer=SGD batch=32 lr0=0.02 name=test_VisDrone_FLDet-N patience=0 epochs=300 save_json=True mosaic=1.0 copy_paste=1.0 mixup=1.0 close_mixup=225 close_mosaic=150 close_copy_paste=75 decay_aug=True > test_VisDrone_FLDet-N.log 2>&1 &

% FLDet-S
yolo detect train data=VisDrone_test.yaml model=FLDet-S.yaml imgsz=640 device=0,1,2,3 optimizer=SGD batch=32 lr0=0.02 name=test_VisDrone_FLDet-S patience=0 epochs=300 save_json=True mosaic=1.0 copy_paste=1.0 mixup=1.0 close_mixup=225 close_mosaic=150 close_copy_paste=75 decay_aug=True > test_VisDrone_FLDet-S.log 2>&1 &

More super parameters about training please refer to Ultralytics YOLOv8 Docs.

2. UAVDT

% FLDet-N
yolo detect train data=UAVDT.yaml model=FLDet-N.yaml imgsz=640 device=0,1,2,3 optimizer=SGD lr0=0.08 name=test_UAVDT_FLDet-N epochs=100 batch=32 save_json=True decay_aug=True mosaic=1.0 copy_paste=1.0 mixup=1.0 close_mixup=75 close_mosaic=50 close_copy_paste=25 > test_UAVDT_FLDet-N.log 2>&1 &

% FLDet-S
yolo detect train data=UAVDT.yaml model=FLDet-S.yaml imgsz=640 device=0,1,2,3 optimizer=SGD lr0=0.08 name=test_UAVDT_FLDet-S epochs=100 batch=32 save_json=True decay_aug=True mosaic=1.0 copy_paste=1.0 mixup=1.0 close_mixup=75 close_mosaic=50 close_copy_paste=25 > test_UAVDT_FLDet-S.log 2>&1 &

Evaluation

yolo detect val data=/path/to/data.yaml model=/path/to/your/best.pt testspeed=False save_json=True name=your-work-dir half=True > val.log 2>&1 &