CLAD: A Contrastive Learning based Approach for Background Debiasing

TL;DR

This paper presents CLAD, a contrastive learning method that reduces background bias in CNNs, improving robustness and generalization by focusing on object foregrounds and effectively sampling negatives.

Contribution

Introduces a novel contrastive learning approach (CLAD) that mitigates background bias in CNNs, with an efficient negative sampling strategy and state-of-the-art results.

Findings

Outperforms previous methods on the Background Challenge dataset by 4.1%.

Enhances CNN robustness by reducing reliance on background features.

Demonstrates effectiveness in debiasing spurious background and texture features.

Abstract

Convolutional neural networks (CNNs) have achieved superhuman performance in multiple vision tasks, especially image classification. However, unlike humans, CNNs leverage spurious features, such as background information to make decisions. This tendency creates different problems in terms of robustness or weak generalization performance. Through our work, we introduce a contrastive learning-based approach (CLAD) to mitigate the background bias in CNNs. CLAD encourages semantic focus on object foregrounds and penalizes learning features from irrelavant backgrounds. Our method also introduces an efficient way of sampling negative samples. We achieve state-of-the-art results on the Background Challenge dataset, outperforming the previous benchmark with a margin of 4.1\%. Our paper shows how CLAD serves as a proof of concept for debiasing of spurious features, such as background and texture (in supplementary material).