Link to paper

• Environment Setup
• Quick Demo
  • Website Demo
  • Code Demo
• Supported models for CLIPCAM
• CAM Variations
• Dataset Preparation
• Evaluation
  • Grid-view Zero-shot Object Localization
    • OpenImage
  • Grid-view Zero-shot Action Localization
    • HICO-DET
  • Single-image Zero-shot Object Localization
    • OpenImage
    • ILSVRC
    • COSMOS
    • Custom Images
• Other features
  • Iterative Mask
  • Weather Attacks

1. create conda enviroment with Python=3.7
   conda create -n clipcam python=3.7
conda activate clipcam
2. install pytorch 1.9.0, torchvision 0.10.0 with compatible cuda version (or any compatible torch version)
   pip install torch==1.9.0 cu111 torchvision==0.10.0 cu111 -f https://download.pytorch.org/whl/torch_stable.html
3. install required package
   pip install -r requirements.txt

Please go to this link for a quick demo.

*P.S. First time user: please follow the instruction on top of the demo website to allow your browser connecting to my server.

python clipcam.py \ 
    --image_path "{single image path or grid image directory (4 images)}" \ 
    --sentence "{input sentence}" \  
    --gpu_id 0 \ 
    --clip_model_name "ViT-B/16" \ 
    --cam_model_name "GradCAM" \ 

example: --clip_model_name ViT-B/16

• ViT-B/16
• ViT-B/32
• RN50
• RN101
• RN50x4
• RN50x16

example: --clip_model_name ViT-B/16-pretrained

• ViT-B/16-pretrained
• ViT-B/32-pretrained
• RN50-pretrained
• RN101-pretrained
• DeiT-pretrained

example: --cam_model_name GradCAM

• GradCAM
• GradCAMPlusPlus
• XGradCAM
• ScoreCAM
• EigenCAM
• EigenGradCAM
• GuidedBackpropReLUModel
• LayerCAM

example: --cam_model_name GradCAM_original

• GradCAM_original
• GradCAMPlusPlus_original
• XGradCAM_original
• ScoreCAM_original
• EigenGradCAM_original
• EigenCAM_original
• GuidedBackpropReLUModel_original
• LayerCAM_original

1. OpenImage V6
   Download OpenImage V6 validation set with data_prep/openimage.py.
2. HICO-DET
   Download HICO-DET from this link.
3. ILSVRC (optional)
   Download ILSVRC validation set.
4. COSMOS (optional)
   Download COSMOS validation set.

1. Dataset structure (OpenImage)

   |--OpenImage
       |--validation
           |--data
               |--{image_path_1}
               |--{image_path_2}
               |-- ...
           |--labels
               |--detections.csv
           |--metadata
               |--classes.csv
   

2. Run evaluate_grid_openimage.py with any model selection

   python evaluate_grid_openimage.py \
       --data_dir Dataset/OpenImage/validation \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/grid/openimage/vitb32-grad' \
       --mask_threshold 0.2 \
       --sentence_prefix 'a photo of ' \
       --attack 'None' \
       --save_result 1
   

1. Dataset structure (HICO-DET)

   |--HICO-DET
       |--images
           |--test
               |--{image_path_1}
               |--{image_path_2}
               |-- ...
           |--train
               |--{image_path_1}
               |--{image_path_2}
               |-- ...
       |--anno.mat
       |--anno_bbox.mat
   

2. Run verb_grid.py for pre-trained model
   Train the model with half of the classes in HICO-DET.
   Or download the fine-tuned checkpoints from this OneDrive.
   --train_mode with 'full', 'few' and 'half' specifies the setting when loading the classes in HICO-DET dataset.

   python verb_grid.py \
       --data_dir datasets/hico-det \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32-pretrained' \
       --cam_model_name 'GradCAM_original' \
       --save_dir 'eval_result/grid/hicodet/vitb32-pretrained-grad' \
       --mask_threshold 0.2 \
       --train_mode 'half' \
       --model_name checkpoints/models/vitb32-pretrained-half-1e-6.pth \ 
       --save_result 1
   

3. Run verb_grid.py for CLIPCAM

   python verb_grid.py \
       --data_dir dataset/hico-det \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/grid/hicodet/vitb32-grad' \
       --mask_threshold 0.2 \
       --save_result 1
   

1. OpenImage
   a. Run evaluate_openimage.py

   python evaluate_openimage.py \
       --data_dir datasets/OpenImage/validation \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/single/openimage/vitb32-grad' \
       --save_result 1 \
       --sentence_prefix 'a photo of ' \
       --distill_num 0 \
       --attack 'None'
   

2. ILSVRC
   
   a. Dataset Structure

   |--ImageNet
       |--validation
           |--{label_1}\
               |--{image_path_1}
               |--{image_path_2}
               |-- ...
           |--{label_2}
           |-- ...
       |--bbox
           |--val
               |--{image_path_1}.xml
               |--{image_path_1}.xml
               |-- ...
   

   b. Run evaluate_imagenet.py

   python evaluate_imagenet.py \
       --data_dir dataset/ImageNet/validation \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/single/imagenet/vitb32-grad' \
       --batch 128 \
       --save_result 1 \
       --sentence_prefix 'sentence' \
       --attack 'None'
   

3. COSMOS, OpenImage and custom images
   
   a. Run evaluate.py with --dataset cosmos or --dataset openimage

   python evaluate.py \
       --data_dir datasets/COSMOS/val \
       --gpu_id 0 \
       --dataset cosmos \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/single/cosmos/vitb32-grad' \
       --distill_num 0 \
       --attack 'None'
   

4. Test on images with custom guiding text a. Put the images in a folder and run evaluate.py b. Run evaluate.py without specifying --dataset

   python evaluate.py \
       --data_dir {path to folder} \
       --gpu_id 0 \
       --clip_model_name 'ViT-B/32' \
       --cam_model_name 'GradCAM' \
       --save_dir 'eval_result/custom-input-vitb32-grad'  \
       --distill_num 0
   

We propose an iterative refinement method based on masking out high neural importance areas to expand attention or enhance weak response regions.

Set --distill_num {n} to iteratively mask out {n} times.

We experimented the ability of CLIPCAM to handle attacked images.

Set --attack fog or --attack snow to create fog or snow attack.

If you find the paper or the code useful for your study, please consider citing the CLIPCAM paper:

@article{clipcam_hsia_icassp2022,
    author = {Hsia, Hsuan-An and Lin, Che-Hsien and Kung, Bo-Han and Chen, Jhao-Ting and Tan, Daniel Stanley and Chen, Jun-Cheng and Hua, Kai-Lung},
    title = "{CLIPCAM: A Simple Baseline for Zero-shot Text-guided Object and Action Localization}",
    booktitle = {Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
    year = {2022}
}

If you have questions regarding the paper or code, please open an issue or email us: Jhao-Ting Chen or Che-Hsien Lin. We will get back to you as soon as possible.