Self-Supervised Object Localization with Joint Graph Partition

Object localization aims to generate a tight bounding box for the target object, which is challenging problem that has been deeply studied in recent years. Since collecting bounding-box labels time-consuming and laborious, many researchers focus on weakly supervised object (WSOL). As appealing self-supervised learning technique shows its powerful function visual tasks, this paper, we take early attempt explore unsupervised by self-supervision. Specifically, adopt different geometric transformations image utilize their parameters as pseudo learning. Then, class-agnostic activation map (CAAM) used highlight potential regions. However, such attention maps merely most discriminative part of objects, will affect quality predicted box. Based motivation same should be equivariant, further design siamese network encodes paired images propose joint graph cluster partition mechanism an manner enhance co-occurrent To validate effectiveness proposed method, extensive experiments are conducted CUB-200-2011, Stanford Cars FGVC-Aircraft datasets. Experimental results show our method outperforms state-of-the-art methods using level supervision, even some weakly-supervised methods.