Gradient-based refined class activation map for weakly supervised object localization
作者：
Highlights：
• We propose a novel Gradient-based Refined CAM approach based on the thought of gradients to mine entire object regions. Our GRCAM makes improvements during the testing stage and does not increase huge training resources.
• We exploit the gradients of the classification loss function to mine the inter-class relationship among the predicted probabilities. The class-specific mask is generated based on inter-class relations to enhance the information of the target class.
• We design a regression function containing the intra-class relationship. The gradients of the regression function are utilized to mine the category consistency for revising the bounding box.
摘要
•We propose a novel Gradient-based Refined CAM approach based on the thought of gradients to mine entire object regions. Our GRCAM makes improvements during the testing stage and does not increase huge training resources.•We exploit the gradients of the classification loss function to mine the inter-class relationship among the predicted probabilities. The class-specific mask is generated based on inter-class relations to enhance the information of the target class.•We design a regression function containing the intra-class relationship. The gradients of the regression function are utilized to mine the category consistency for revising the bounding box.
论文关键词：Weakly supervised object localization,Gradients of loss function,Class-specific mask,Bounding box revision,Category consistency
论文评审过程：Received 13 February 2021, Revised 2 March 2022, Accepted 21 March 2022, Available online 22 March 2022, Version of Record 30 March 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108664
