PyTorch Implementation of Spatial Sensitive Grad-CAM &
Spatial Sensitive Grad-CAM++

Overview

Unofficial PyTorch implementation of Spatial Sensitive Grad-CAM (SSGrad-CAM) and Spatial Sensitive Grad-CAM++ (SSGrad-CAM++) for generating instance-specific heat maps for object detectors.
This repository provides minimal reference code to generate heat maps.

For implementation, we heavily refer to the VX-CODE implementation (VX-CODE) and base our code on Detectron2. For DETR, we additionally refer to the official DETR repository.

What this repository provides

• generate_heatmap.py: CLI to run SSGrad-CAM / SSGrad-CAM++ and save heatmaps
• Sample runner scripts for Faster R-CNN and DETR (run_on_faster.sh, run_on_detr.sh)
• Templates for configs (configs/), model weights (models/), and dataset references (datasets/)
• Helper script to download experiment weights: weights_dl.sh
• Helper script to download MS-COCO 2017 data: coco_dl.sh

Quick start

If the detectron2 submodule was not fetched at clone time, initialize/update it first: git submodule update --init --recursive

1. Set up a Python environment and install dependencies (PyTorch, etc.).
2. Download weights with bash weights_dl.sh and place them under models/ if needed. (The Google Drive files are the official weights provided by the original repositories.)
3. Download MS-COCO data (e.g., 2017 val images/annotations) into datasets/ with the expected detectron2 structure. You can use bash coco_dl.sh to fetch val2017 + annotations by default. If COCO is not available, the scripts can still run on the sample images in datasets/test_imgs.
4. Run examples: bash run_on_faster.sh or bash run_on_detr.sh. Inside each script, switch METHOD/OUTPUTDIR between ssgradcam and ssgradcampp to choose SSGrad-CAM vs. SSGrad-CAM++. Output directories are set inside each script via --output-dir.
5. For direct CLI usage, check options with python generate_heatmap.py --help.

Result

Example heatmaps generated by running generate_heatmap.py. Warmer colors highlight regions that contribute most to the detector’s prediction.

Reference

For implementing SSGrad-CAM, we refer to the following ICIP paper and its extended journal version published in IEEE Access.

@INPROCEEDINGS{9897350,
  author={Yamauchi, Toshinori and Ishikawa, Masayoshi},
  booktitle={2022 IEEE International Conference on Image Processing (ICIP)}, 
  title={Spatial Sensitive GRAD-CAM: Visual Explanations for Object Detection by Incorporating Spatial Sensitivity}, 
  year={2022},
  volume={},
  number={},
  pages={256-260},
  doi={10.1109/ICIP46576.2022.9897350}}

@ARTICLE{11270886,
  author={Yamauchi, Toshinori},
  journal={IEEE Access}, 
  title={Spatial Sensitive Grad-CAM: Toward Instance-Specific Explanations for Object Detectors by Incorporating Spatial Sensitivity}, 
  year={2025},
  volume={13},
  number={},
  pages={202086-202102},
  doi={10.1109/ACCESS.2025.3637887}}

For implementing SSGrad-CAM++, we refer to the following CVPR workshop paper and its extended journal version published in CVIU.

@InProceedings{Yamauchi_2024_CVPR,
    author    = {Yamauchi, Toshinori},
    title     = {Spatial Sensitive Grad-CAM++: Improved Visual Explanation for Object Detectors via Weighted Combination of Gradient Map},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2024},
    pages     = {8164-8168}
}

@article{YAMAUCHI2026104658,
title = {Spatial Sensitive Grad-CAM++: Towards High-Quality Visual Explanations for Object Detectors via Weighted Combination of Gradient Maps},
journal = {Computer Vision and Image Understanding},
volume = {264},
pages = {104658},
year = {2026},
issn = {1077-3142},
doi = {https://doi.org/10.1016/j.cviu.2026.104658},
url = {https://www.sciencedirect.com/science/article/pii/S1077314226000251},
author = {Toshinori Yamauchi},
}

Paper links

Spatial Sensitive Grad-CAM (SSGrad-CAM)

• ICIP 2022: https://ieeexplore.ieee.org/document/9897350
• IEEE Access (extended version): https://ieeexplore.ieee.org/document/11270886

Spatial Sensitive Grad-CAM++ (SSGrad-CAM++)

• CVPR 2024 Workshop: https://openaccess.thecvf.com/content/CVPR2024W/XAI4CV/html/Yamauchi_Spatial_Sensitive_Grad-CAM_Improved_Visual_Explanation_for_Object_Detectors_via_CVPRW_2024_paper.html
• CVIU (extended version): https://www.sciencedirect.com/science/article/pii/S1077314226000251