Object representation enhancement for self‐supervised colocalization

Abstract

Self-supervised colocalization is to localize common objects in the data set containing only one superclass without using human-annotated labels. Existing methods achieve impressive results by employing self-supervised pretext learning. However, a common limitation still exists. They either tend to overextend activations to the

A Task of Convergent AI Ethics Education in School Curriculum - with emphasis on the major of AI Convergent Education in graduate schools of education and the class of ‘AI Ethics’ -

“…We report the performances of our method and the state-of-the-art methods in terms of GT-known Loc on five commonly used datasets in Table 1. We compare our method with recent self-supervised object localization methods including ORE [29], PsyNet [1], Ki et al [23], JGP [38] and C 2 AM [48], as well as object localization methods without finetuning including DDT [45], MO [54], LOST [37], and TokeneCut [42]. As shown in Table 1, our method significantly surpasses other methods on all benchmark datasets.…”

“…These features are extracted from regions of the image that contain rich information to distinguish them from other instances, and these regions are more likely to be associated with significant objects. Likewise, prior self-supervised colocalization studies [1,23,29,38] also use the magnitude of feature vectors as a clue to discover object regions.…”

“…Beyond learning with less supervision, most recent works [45,44,54,1,23,29,38,48,37,42] focus on the object localization task using self-supervised or unsupervised learning that does not require any human annotated labels. These works address the problem of identifying which regions are more likely to contain the foreground object, which is a salient object in an image.…”

“…These works address the problem of identifying which regions are more likely to contain the foreground object, which is a salient object in an image. In order to discover foreground object regions, several methods [1,23,29,38] attempt to use the magnitude of feature vectors as a clue for class-agnostic activation maps (CAAM) [1]. Most of these methods rely heavily on pre-trained models designed for the image classification task.…”

“…Class activation maps (CAM) [59] are frequently used to provide visually interpretable information about a specific class the model has learned, but a detailed explanation of how the model makes its predictions arXiv:2309.04172v1 [cs.CV] 8 Sep 2023 remains unclear. This limitation becomes even more severe for CAAM, which is commonly used in self-supervised object localization methods [1,23,29,38].…”