Weakly Supervised Object Localization Based On Shallow Features

Weakly supervised object localization utilizes coarse-grained annotations to achieve the object localization task.In the paper,it studies the issues in the existing weakly supervised object localization methods,such as locating local regions,blurring edge information,and ignoring the category consistency.The main research achievements include:(1)An Attention-based Multi-branches Network for weakly supervised object localization is proposed.Aiming to the problem that weakly supervised localization algorithms only identify the local areas,the multi-branch structure constructs parallel branches to mine the edge regions of the target object by cascading the shallow feature maps in the classification network.In addition,the proposed channel-attention mechanism assigns appropriate attention for different channels to preserve the non-discriminative areas.Extensive experiments on the ILSVRC dataset show the proposed method achieves excellent localization performance and the top-1 localization error achieves 46.97%.(2)A Dual-Edges Localization framework for weakly supervised object localization is proposed.Although the multi-branches network enables to mine the edge information,it is not ingenious to exchange the increase of training resources for the improvement of localization accuracy.In order to alleviate the problem,the proposed method constructs the similarity measurement to extract the shallow edge features for enhancing the edge information of the target object in the last convolutional feature map in an offline way during the testing stage.Extensive experiments on the ILSVRC dataset show that the proposed method achieves a better localization performance with less training resources,and the top-1 localization error achieves 45.89%.(3)A Dual-Edges Localization framework based on category consistency for weakly supervised object localization is proposed.The above algorithms mine the edge information of the target object,while the category consistency is not considered.Thus,based on the dual-edges localization framework,the proposed method introduces category consistency information to enhance the ability of mining intra-class information.Extensive experiments on the ILSVRC dataset show that the introduction of category consistency information can achieve better localization,and the top-1 localization error achieves 45.56%.