Download the pretrained parameters (e.g., moco and detco) at here and put them in the current directory.
├── WSSS/
| ├── core
| ├—— ...
| ├—— moco_r50_v2-e3b0c442.pth
| └── detco_200ep.pth
- Let's take the Market1501 dataset as an example. Make sure your
data/Market1501folder is structured as follows:
├── bounding_box_train/
| ├── xxx.jpg
| ├── ...
├── bounding_box_test/
└── query/
- Then specify the input size in
train_CCAM.py
train_transform = transforms.Compose([
# the input size is decided by the adopted datasets
transforms.Resize(size=(256, 128)),
transforms.RandomHorizontalFlip(),
# transforms.RandomCrop(size=(448, 448)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))
])
- To train CCAM on the custom dataset
OMP_NUM_THREADS=16 CUDA_VISIBLE_DEVICES=5 python train_CCAM.py --tag CCAM_Maket1501_MOCO --batch_size 128 --pretrained mocov2 --alpha 0.25
The code will create experiment folders for model checkpoints (./experiment/models), log files (.experiments/logs) and visualization (./experiments/images/).
├── experiments/
| ├── checkpoints
| ├—— images
If you are using our code, please consider citing our paper.
@InProceedings{Xie_2022_CVPR,
author = {Xie, Jinheng and Xiang, Jianfeng and Chen, Junliang and Hou, Xianxu and Zhao, Xiaodong and Shen, Linlin},
title = {C2AM: Contrastive Learning of Class-Agnostic Activation Map for Weakly Supervised Object Localization and Semantic Segmentation},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {989-998}
}
@article{xie2022contrastive,
title={Contrastive learning of Class-agnostic Activation Map for Weakly Supervised Object Localization and Semantic Segmentation},
author={Xie, Jinheng and Xiang, Jianfeng and Chen, Junliang and Hou, Xianxu and Zhao, Xiaodong and Shen, Linlin},
journal={arXiv preprint arXiv:2203.13505},
year={2022}
}