lucasdavid / minmax-cam

Official implementation for the paper "MinMax-CAM: Improving Focus of CAM-Based Visualization Techniques in Multi-Label Problems"

MinMax-CAM

Official implementation for the paper "MinMax-CAM: Improving Focus of CAM-Based Visualization Techniques in Multi-Label Problems", presented at 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISAPP 2022).

Implementations:

• MinMax-CAM
• D-MinMax-CAM
• MinMax-Grad-CAM
• D-MinMax-Grad-CAM
• Kernel Usage Regularization

Examples

The figure below illustrates the explaining regions proposed by multiple visual explaining techniques, for each one of the classified labels in each image. The techniques are, from left to right, CAM, Grad-CAM++, Score-CAM, MinMax-CAM and D-MinMax-CAM.

Summary

We list below the notebooks used to train, evaluate and observe the multiple CAM-based visualization methods studied in our work.

Training

#	Dataset	Name	Description
1	VOC 2007	resnet101-multilabel-pascal-voc-2007	training notebook for a ResNet101 multi-label classifier over the Pascal VOC 2007 dataset
2	VOC 2007	vgg16-multilabel-pascal-voc-2007	training notebook for a VGG16-GAP multi-label classifier over the Pascal VOC 2007 dataset
3	VOC 2007	efficientnetb6-multilabel-pascal-voc-2007	training notebook for a EfficientNet-B6 multi-label classifier over the Pascal VOC 2007 dataset
4	VOC 2012	resnet101-multilabel-pascal-voc-2012	training notebook for a ResNet101 multi-label classifier over the Pascal VOC 2012 dataset
5	COCO 2017	resnet101-multilabel-pascal-coco	training notebook for a ResNet101 multi-label classifier over the COCO 2017 dataset
6	P:AfS	resnet101-multilabel-amazon-from-space	training notebook for a ResNet101 multi-label classifier over the Planet: Understanding the Amazon from Space dataset

Evaluation

#	Dataset	Name	Description
1	VOC 2007	cam-benchmarks-voc-2007	evaluation notebook for multiple explaining methods, using the VGG16-GAP, ResNet101 and EfficientNet-B6 networks over the Pascal VOC 2007 dataset
2	VOC 2012	cam-benchmarks-voc-2012	evaluation notebook for multiple explaining methods, using the ResNet101 network over the Pascal VOC 2012 dataset
3	COCO 2017	cam-benchmarks-coco	evaluation notebook for multiple explaining methods, using the ResNet101 network over the COCO 2017 dataset
4	P:AfS	cam-benchmarks-amazon-from-space	evaluation notebook for multiple explaining methods, using the ResNet101 network over the Planet: Understanding the Amazon from Space dataset

Analysis

#	Dataset	Name	Description
1	VOC 2007	study-of-sigmoid-multilabel-voc-2007	analysis notebook for multiple explaining methods, using the ResNet101 network over the Pascal VOC 2007 dataset
2	VOC 2012	study-of-sigmoid-multilabel-voc-2012	analysis notebook for multiple explaining methods, using the ResNet101 network over the Pascal VOC 2012 dataset
3	P:AfS	study-of-sigmoid-multilabel-amazon-from-space	analysis notebook for multiple explaining methods, using the ResNet101 network over the Planet: Understanding the Amazon from Space dataset

Citation

@conference{visapp22,
author={Lucas David. and Helio Pedrini. and Zanoni Dias.},
title={MinMax-CAM: Improving Focus of CAM-based Visualization Techniques in Multi-label Problems},
booktitle={Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP,},
year={2022},
pages={106-117},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010807800003124},
isbn={978-989-758-555-5},
}