"TDAM: Top-Down Attention Module for Contextually Guided Feature Selection in CNNs"

• PyTorch implementation for paper "Top-Down Attention Module for Contextually Guided Feature Selection in CNNs" (ECCV 2022; paper).

• To run code, ideally create a virtual/conda environment and install requirements listed in requirements.txt by running:

pip install -r requirements.txt

• For module usage and performing training/analysis, please see provided scripts in training_and_analysis_scripts directory (specifically TDAM_usage_and_visualization.ipynb with instructions in that directory's README.md.

• For just the module and model integration/implementation code, please see modules_and_models directory.

ImageNet-1k pre-trained models

Model	Top-1(%)	Top-5(%)	GoogleDrive
TDAM(t2,m2)-RNet18	72.16	90.61	TD_ResNet18
TDAM(t2,m2)-RNet34	75.75	92.58	TD_ResNet34
TDAM(t2,m1)-RNet50	78.96	94.19	TD_ResNet50
TDAM(t2,m1)-RNet101	81.62	95.76	TD_ResNet101

Citation

@inproceedings{jaiswal2022tdam,
 title={TDAM: Top-Down Attention Module for Contextually Guided Feature Selection in CNNs},
 author={Jaiswal, Shantanu and Fernando, Basura and Tan, Cheston},
 booktitle={European Conference on Computer Vision},
 pages={259--276},
 year={2022},
 organization={Springer}
}

Code environment

The codebase and associated experiments are performed in following environment:

• OS: Ubuntu 20.04.4 LTS
• CUDA: 11.4
• GPU: NVIDIA Tesla V100 DGXS (16GB)
• Python: 3.8.10
• Python packages/toolkits: See requirements.txt

Acknowledgement

The codebase utilizes the timm and torchvision libraries.

License

This project's codebase is released under the MIT license. Please see the LICENSE file for more information.

Contact Information

In case of any suggestions or questions, please leave a message here or contact me directly at jaiswals@ihpc.a-star.edu.sg, thanks!