Keras Custom Class Activation Map Implementation (multi CAM overlay options)

Class Activation Map is an unsupervised way of doing object localization with accuracy near par with supervised methods.

Custom implementation of CAM using Keras with a Tensorflow backend. Takes in an image and returns

1. Prediction accuracy (alongside class label and heatmap legend)
2. Heatmap showing which parts of the image constituted the majority of the classification. There is the option to overlay a single or multiple CAM heatmap(s) on top of the original image.

Getting Started

Within CAM.py, please read docstring of:

1. get_cam
2. get_multi_layered_cam
3. overlay_prediction_on_image (writes predictions on top of CAM)

Must have a keras model with a global average pooling layer after the final convolution layer followed by a single input -> output layer (Part 1 jupyter notebook)

Part 2 jupyer notebook contains examples of what CAM.py can do

Prerequisites

pip install matplotlib
pip install keras
pip install numpy

OPTIONAL: OpenCV for CAM overlay on image

Results

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

Original paper: https://arxiv.org/pdf/1512.04150.pdf