{"title":"DHS-DETR: Efficient DETRs with dynamic head switching","authors":"","doi":"10.1016/j.cviu.2024.104106","DOIUrl":null,"url":null,"abstract":"<div><p>Detection Transformer (DETR) and its variants have emerged a new paradigm to object detection, but their high computational cost hinders practical applications. By investigating their essential components, we found that the transformer-based head usually occupies a significant amount of computation. Through further comparing heavy and light transformer heads, we observed that both heads produced satisfactory results for easy images while showing a noticeable difference for hard images. Inspired by these findings, we propose a dynamic head switching (DHS) strategy to dynamically select the proper head for each image at inference for a better balance of efficiency and accuracy. Specifically, our DETR model incorporates multiple heads with different computational complexity and a lightweight module which selects proper heads for given images. This module is optimized to maximize detection accuracy while adhering to the overall computational budget limitations. To minimize the potential accuracy drop when executing the lighter heads, we propose online head distillation (OHD) to improve the accuracy of the lighter heads with the help of the heavier head. Extensive experiments on the MS COCO dataset validated the effectiveness of the proposed method, which demonstrated a better accuracy–efficiency trade-off compared to the baseline using static heads.</p></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":null,"pages":null},"PeriodicalIF":4.3000,"publicationDate":"2024-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Computer Vision and Image Understanding","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1077314224001875","RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Detection Transformer (DETR) and its variants have emerged a new paradigm to object detection, but their high computational cost hinders practical applications. By investigating their essential components, we found that the transformer-based head usually occupies a significant amount of computation. Through further comparing heavy and light transformer heads, we observed that both heads produced satisfactory results for easy images while showing a noticeable difference for hard images. Inspired by these findings, we propose a dynamic head switching (DHS) strategy to dynamically select the proper head for each image at inference for a better balance of efficiency and accuracy. Specifically, our DETR model incorporates multiple heads with different computational complexity and a lightweight module which selects proper heads for given images. This module is optimized to maximize detection accuracy while adhering to the overall computational budget limitations. To minimize the potential accuracy drop when executing the lighter heads, we propose online head distillation (OHD) to improve the accuracy of the lighter heads with the help of the heavier head. Extensive experiments on the MS COCO dataset validated the effectiveness of the proposed method, which demonstrated a better accuracy–efficiency trade-off compared to the baseline using static heads.
期刊介绍:
The central focus of this journal is the computer analysis of pictorial information. Computer Vision and Image Understanding publishes papers covering all aspects of image analysis from the low-level, iconic processes of early vision to the high-level, symbolic processes of recognition and interpretation. A wide range of topics in the image understanding area is covered, including papers offering insights that differ from predominant views.
Research Areas Include:
• Theory
• Early vision
• Data structures and representations
• Shape
• Range
• Motion
• Matching and recognition
• Architecture and languages
• Vision systems