Optimizing Relevance Maps of Vision Transformers Improves Robustness
This addresses robustness issues in visual classification models, particularly for domain shifts, but is incremental as it builds on existing Vision Transformer methods.
The paper tackles the problem of vision models relying on image background instead of foreground, which reduces robustness to distribution shifts, by proposing a fine-tuning method that manipulates relevance maps to focus on foreground objects, resulting in marked improvement in robustness to domain shifts.
It has been observed that visual classification models often rely mostly on the image background, neglecting the foreground, which hurts their robustness to distribution changes. To alleviate this shortcoming, we propose to monitor the model's relevancy signal and manipulate it such that the model is focused on the foreground object. This is done as a finetuning step, involving relatively few samples consisting of pairs of images and their associated foreground masks. Specifically, we encourage the model's relevancy map (i) to assign lower relevance to background regions, (ii) to consider as much information as possible from the foreground, and (iii) we encourage the decisions to have high confidence. When applied to Vision Transformer (ViT) models, a marked improvement in robustness to domain shifts is observed. Moreover, the foreground masks can be obtained automatically, from a self-supervised variant of the ViT model itself; therefore no additional supervision is required.