Weakly supervised object co-localization via sharing parts based on a joint Bayesian model

Abstract

Objects in images are characterized by intra-class variation, inter-class diversity, and noisy images. These characteristics pose a challenge to object localization. To address this issue, we present a novel joint Bayesian model for weakly-supervised object localization. The differences compared to previous discriminative methods are evaluated in three aspects: (1) We co-localize the similar object per class through transferring shared parts, which are pooling by modeling object, parts and features within and between-class; (2) Labels are given at class level to provide strong supervision for features and corresponding parts; (3) Noisy images are considered by leveraging a constraint on the detection of shared parts. In addition, our methods are evaluated by extensive experiments. The results indicated outperformance of the state-of-the-art approaches with almost 7% and 1.5% improvements in comparison to the previous methods on PASCAL VOC 2007 6 × 2 and Object Discovery datasets, respectively.