won-bae / rethinkingCAM

Official implementation of Rethinking Class Activation Mapping for Weakly Supervised Object Localization (ECCV 2020)

Rethinking Class Activation Mapping for Weakly Supervised Object Localization [Paper][Project Page]

Datasets

Create symlinks for datasets: CUB-200-2011 and ImageNet-1k. Note that the name of the directory for CUB should be 'CUB_200_2011'.

cd /path/to/this/repo
mkdir data && cd data
ln -s /path/to/cub/dataset
ln -s /path/to/imagenet/dataset

Setup

Setup the pipeline by installing dependencies, pretrained models and utils.

cd /path/to/this/repo
bash scripts/setup.sh --data_setup

Train

To train a model, use the command below. Tag can be an arbitrary string.

cd /path/to/this/repo
bash scripts/run_train.sh --config_path configs/vgg16_tap.yml --tag vgg16_tap --train_dir /path/to/train/dir

Eval

To eval a model, use the command below. Note that checkpoint dir has to be synchronized with train_dir used for training.

cd /path/to/this/repo
bash scripts/run_eval.sh --config_path configs/baseline/vgg16_tap.yml --tag vgg16_tap --checkpoint_dir /path/to/checkpoint/dir

NOTE: NWC and PaS can be used in evaluation time. In configs/vgg16_tap.yml, NWC and PaS can be employed simply changing paraameters under 'eval'. For NWC, change 'truncate' to True. For PaS, change 'percentile' to 90 and 'loc_threshold' to 0.35.

TODO

Model and configs for the other backbones such as Resnet50-SE, Inception need to be added.

Citation

If you use this code or model for your research, please cite:

@InProceedings{bae2020rethinkingCAM,
  author = {Wonho Bae and Junhyug Noh and Gunhee Kim},
  title = {Rethinking Class Activation Mapping for Weakly Supervised Object Localization},
  booktitle = {The European Conference on Computer Vision (ECCV)},
  month = {August},
  year = {2020}
}

Acknowledgment

Many of core functions are borrowed from ADL.