djamahl99 / underwater-cots-da

Edge Deployable Online Domain Adaptation for Underwater Object Detection

Edge Deployable Online Domain Adaptation for Underwater Object Detection

Results and Models

All evaluated on ReefScan$^{\operatorname{pos}}_{\operatorname{test}}$.

Technique	Adaptation Set	F2 0.3:0.8	Download
YOLOV5 Source	NA	$32.54%$	...
YOLOV5 Relighting	Source and ReefScan$^{\operatorname{pos}}_{\operatorname{train}}$	$39.99%$	Lightnet
YOLOV5 Combined	ReefScan$^{\operatorname{pos}}_{\operatorname{train}}$	$54.8%$	ckpt
YOLOV5 Combined	ReefScan$^{\operatorname{sub}}_{\operatorname{train}}$	$52.3%$	ckpt

Darknet is available.

Requirements

• pytorch
• mmyolo

Datasets

Kaggle: Kaggle COTS dataset

AIMS: AIMS COTS dataset

Splits

Google Drive.

Testing

AIMS

 python evaluate.py --dataset aims_sep

Kaggle

python evaluate.py --dataset kaggle

Training

Kaggle -> AIMS

 python train.py --run-name aims-adaptation-yolov5 --model yolov5 --batch-size 2 --teacher-score-thresh 0.5
 python train.py --run-name aims-adaptation-yolov8 --model yolov8 --batch-size 1 --teacher-score-thresh 0.5

Kaggle -> AIMS multiday

 python train_multiday.py --run-name aims-adaptation-yolov5 --model yolov5 --batch-size 2 --teacher-score-thresh 0.7

Relighting AIMS -> Kaggle

 python train_lightnet.py --run-name aims-relighting --batch-size 2

Acknowledgments

The code is based on DANNet, DINO and mmyolo.

Contact

• Djamahl Etchegaray ([e-mail])