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

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

1. Install other requirements

shell pip install -e .

Usage

Data preparation

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

Training

1. VisDrone2019

```shell % FLDet-N yolo detect train data=VisDronetest.yaml model=FLDet-N.yaml imgsz=640 device=0,1,2,3 optimizer=SGD batch=32 lr0=0.02 name=testVisDroneFLDet-N patience=0 epochs=300 savejson=True mosaic=1.0 copypaste=1.0 mixup=1.0 closemixup=225 closemosaic=150 closecopypaste=75 decayaug=True > testVisDroneFLDet-N.log 2>&1 &

% FLDet-S yolo detect train data=VisDronetest.yaml model=FLDet-S.yaml imgsz=640 device=0,1,2,3 optimizer=SGD batch=32 lr0=0.02 name=testVisDroneFLDet-S patience=0 epochs=300 savejson=True mosaic=1.0 copypaste=1.0 mixup=1.0 closemixup=225 closemosaic=150 closecopypaste=75 decayaug=True > testVisDroneFLDet-S.log 2>&1 & ```

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

2. UAVDT

```shell % 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=testUAVDTFLDet-N epochs=100 batch=32 savejson=True decayaug=True mosaic=1.0 copypaste=1.0 mixup=1.0 closemixup=75 closemosaic=50 closecopypaste=25 > testUAVDT_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=testUAVDTFLDet-S epochs=100 batch=32 savejson=True decayaug=True mosaic=1.0 copypaste=1.0 mixup=1.0 closemixup=75 closemosaic=50 closecopypaste=25 > testUAVDT_FLDet-S.log 2>&1 & ```

Evaluation

shell 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 &