WSCUOD

The code for "Weakly-supervised Contrastive Learning for Unsupervised Object Discovery" The checkpoint for evaluation could be downloaded from wcl-16-final.pth

The results for object segmentation in VOC2007, VOC2012, COCO20K, DUTS-Test, DUT-OMRON and ECSSD could be downloaded from seg_results

1. Dependencies

This code was implemented with Python 3.8, PyTorch 1.9.1+cu111 and CUDA 11.6.

2. Data

The dataset for segmentation could be downloaded from ECSSD, DUTS and DUT-OMRON; The dataset for detection could be downloaded from VOC2007, VOC2012 and COCO20K

3. Training

To train the model, please check the dataset directory in data/dataloader.py (image_folder_prefix) and the pre-trained model directory in wcl.py (--pretrained_path) and run:

bash run_pipeline.sh

4. Testing and Evaluation

Please check:

the testing dataset directory in ddt_dino_sig.py (--test_root and test_dataset)
the checkpoint directory in ddt_dino_sig.py (--pretr_path):
the testing dataset directory and save path in measure.py(pred_dir0, test_datasets) and run

python ddt_dino_sig.py
python measure.py

5. Acknowledgement

We acknowledge these excellent works that inspire our project: Weakly Contrastive Learning, DDT and DINO.