{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/optimizing-relevance-maps-of-vision","title":"Optimizing Relevance Maps of Vision Transformers Improves Robustness","arxiv_id":"2206.01161","date":"2022-06-02","proceeding":null,"authors":["Hila Chefer","Idan Schwartz","Lior Wolf"],"abstract":"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.","url_abs":"https://arxiv.org/abs/2206.01161v1","url_pdf":"https://arxiv.org/pdf/2206.01161v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"optimizing-relevance-maps-of-vision","repo_url":"https://github.com/hila-chefer/robustvit","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"out-of-distribution-generalization","task_name":"Out-of-Distribution Generalization"}],"methods":[{"method_slug":"absolute-position-encodings","method_name":"Absolute Position Encodings"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"attention","method_name":"Attention"},{"method_slug":"bpe","method_name":"BPE"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"label-smoothing","method_name":"Label Smoothing"},{"method_slug":"layer-normalization","method_name":"Layer Normalization"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"position-wise-feed-forward-layer","method_name":"Position-Wise Feed-Forward Layer"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"transformer","method_name":"Transformer"},{"method_slug":"vision-transformer","method_name":"Vision Transformer"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-L (Opt Relevance)","rank_in_archive_order":24,"of":106,"metrics":{"Top-1 Accuracy":"52.0","Top-5 Accuracy":"73.5"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-B (Opt Relevance)","rank_in_archive_order":32,"of":106,"metrics":{"Top-1 Accuracy":"47.1","Top-5 Accuracy":"70"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-L","rank_in_archive_order":36,"of":106,"metrics":{"Top-1 Accuracy":"46.5","Top-5 Accuracy":"68.3"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"ViT-L (Opt Relevance)","rank_in_archive_order":37,"of":106,"metrics":{"Top-1 Accuracy":"43.2","Top-5 Accuracy":"65.8"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"ViT-B (Opt Relevance)","rank_in_archive_order":40,"of":106,"metrics":{"Top-1 Accuracy":"42.2","Top-5 Accuracy":"65.1"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-B","rank_in_archive_order":42,"of":106,"metrics":{"Top-1 Accuracy":"41.4","Top-5 Accuracy":"63.7"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-S (Opt Relevance)","rank_in_archive_order":45,"of":106,"metrics":{"Top-1 Accuracy":"39.3","Top-5 Accuracy":"61.7"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"ViT-L","rank_in_archive_order":48,"of":106,"metrics":{"Top-1 Accuracy":"37.4","Top-5 Accuracy":"59.5"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"DeiT-L (Opt Relevance)","rank_in_archive_order":49,"of":106,"metrics":{"Top-1 Accuracy":"36.3","Top-5 Accuracy":"56.6"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"ViT-B","rank_in_archive_order":54,"of":106,"metrics":{"Top-1 Accuracy":"35.1","Top-5 Accuracy":"56.4"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"AR-S","rank_in_archive_order":56,"of":106,"metrics":{"Top-1 Accuracy":"34.3","Top-5 Accuracy":"55.8"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"DeiT-S (Opt Relevance)","rank_in_archive_order":60,"of":106,"metrics":{"Top-1 Accuracy":"31.6","Top-5 Accuracy":"53"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"DeiT-L","rank_in_archive_order":62,"of":106,"metrics":{"Top-1 Accuracy":"31.4","Top-5 Accuracy":"48.5"},"uses_additional_data":true},{"leaderboard":"/sota/image-classification-on-objectnet","task":"Image Classification","dataset":"ObjectNet","model":"DeiT-S","rank_in_archive_order":76,"of":106,"metrics":{"Top-1 Accuracy":"28.3","Top-5 Accuracy":"47.3"},"uses_additional_data":true}],"syntology":{"syntology_url":"https://syntology.ai/paper/2206.01161","atlas_url":"https://app.syntology.ai/?focus=2206.01161","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}