Grid Saliency for Context Explanations of Semantic Segmentation

TL;DR

This paper introduces grid saliency, a novel method for generating spatially coherent visual explanations for semantic segmentation networks, enabling better understanding of object-context relationships and bias detection.

Contribution

The paper extends saliency methods to dense prediction models, proposing grid saliency for context explanations in semantic segmentation, and demonstrates its effectiveness on synthetic and real datasets.

Findings

Grid saliency provides clear visual explanations of context in segmentation.

It can detect and localize biases in datasets.

Effective on both synthetic and real-world datasets.

Abstract

Recently, there has been a growing interest in developing saliency methods that provide visual explanations of network predictions. Still, the usability of existing methods is limited to image classification models. To overcome this limitation, we extend the existing approaches to generate grid saliencies, which provide spatially coherent visual explanations for (pixel-level) dense prediction networks. As the proposed grid saliency allows to spatially disentangle the object and its context, we specifically explore its potential to produce context explanations for semantic segmentation networks, discovering which context most influences the class predictions inside a target object area. We investigate the effectiveness of grid saliency on a synthetic dataset with an artificially induced bias between objects and their context as well as on the real-world Cityscapes dataset using state-of-the-art segmentation networks. Our results show that grid saliency can be successfully used to provide easily interpretable context explanations and, moreover, can be employed for detecting and localizing contextual biases present in the data.