Hide-CAM: Finding Multiple Discriminative Regions in Weakly Supervised Location

Weakly supervised localization is a more challenging task due to the absence of an object's annotation. Because the depth convolution feature can well represent the spatial information of the object, the position of the object can be located by the saliency study of the image. However, the most discriminative area tends to focus too much on the details of the object and lacks the perception of the object's overall structure, whereas the information of the complementary object regions is the complement of the most discriminative area. The combination of these information types can fully express the global information of the object. Therefore, our method takes into account the entire area of the object rather than the most discriminative area. In this paper, the hide strategy is used to locate the most discriminative and the complementary object regions of the object. First, we use CAM to extract the most discriminative area. Next, we mask the most discriminative area and use CAM to extract complementary object regions in the masked image. Finally, the two areas are integrated to complete the task of location. Our method only needs the classification label of the image instead of a detailed object annotation. The operation is simple and convenient, and does not require training a complex model or additional annotation. Experiments show our method achieves good results in ILSVRC 2012 validation.