We release the code of Self-Supervised Marine Organism Detection from Underwater Images in our papers:
We design an Self-Supervised Marine Organism Detection framework for underwater images to detect organisms. We propose a set of underwater image augmentation strategies to enhance the quality of representation we learned, besides, we also propose a novel Underwater Attention module to explore effective underwater representation for marine organism detection. This code is based on the Simsiam and mmdetection codebase (v2.13.0).
- Linux or macOS (Windows is in experimental support)
-
Python 3.8+
-
PyTorch 1.9+
-
CUDA 10.2+ (If you build PyTorch from source, CUDA 10.0 is also compatible)
-
GCC 5+
-
MMCV
HabCam, TRASH ICRA-2019, UIEBD, RUIE, URPC2021.
-
Contact dataset provider, and download the datasets and annotations.
-
Put all images and annotation files to ./data folder.
- To train a SSL Method on Pretext dataset, you can run the script:
CUDA_VISIBLE_DEVICES=0,1 python DDP_simsiam_ccrop_pretrain.py configs/small/underwater/simsiam_ccrop_pretrain.py
- You can also set the variables (CONFIG_FILE, GPU_NUM) in pretrain.sh and configs/small/underwater/simsiam_ccrop_pretrain.py, and then run the script:
bash pretrain.sh
- After pretext task training, you can run the folllowing code to transfer the weight file to downstream task：
python self-weight_converter_pre.py
- Besides, the downstream task is training on the mmdetection version 2.13.0.
We will provide the models and results later.
We really appreciate the contributors of following codebases.