ERL-Net

Note

This project code is for Underwater Object Detection.

Install

It is based on MMdetection, please refer to install.md to install MMdetection.

Dataset

We conduct experiments on the two challenging underwater datasets UTDAC2020 and Brackish dataset. UTDAC2020 is the newest underwater dataset which is from Underwater Target Detection Algorithm Competition 2020. In addition, there are many wrong annotations in the original dataset, thus we manually corrected the wrong data annotations on UTDAC2020. The refined UTDAC2020 dataset is open-sourced in https://drive.google.com/file/d/1avyB-ht3VxNERHpAwNTuBRFOxiXDMczI/view?usp=sharing.

The structure of this dataset is:

├── data
│   ├── UTDAC2020
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── annotations

Train

Train on UTDAC2020 dataset

(1) Cascade R-CNN

python tools/train.py configs/erl/erl_cascade_utdac.py

(2) Faster R-CNN

python tools/train.py configs/erl/erl_faster_rcnn_utdac.py

(3) RetinaNet

python tools/train.py configs/erl/erl_retina_utdac

Train on Brackish dataset

(1) Cascade R-CNN

python tools/train.py configs/erl/erl_cascade_brackish.py

(2) Faster R-CNN

python tools/train.py configs/erl/erl_faster_rcnn_brackish.py

(3) RetinaNet

python tools/train.py configs/erl/erl_retina_brackish.py

Test and Evaluation

Please follow the steps of MMdetection.

Demo

UTDAC2020 dataset

Brackish dataset