Estimation of Near-Instance-Level Attribute Bottleneck for Zero-Shot Learning

Zero-Shot Learning (ZSL) involves transferring knowledge from seen classes to unseen classes by establishing connections between visual and semantic spaces. Traditional ZSL methods identify novel classes by class-level attribute vectors, which implies an information bottleneck. These approaches often use class-level attribute vectors as the fitting target during training, disregarding the individual variations within a class. Moreover, the attributes used for training lack location information and are prone to mismatch with local regions of visual features. To this end, we introduce a Near-Instance-Level Attribute Bottleneck (IAB) to alter class-level attribute vectors as well as visual features throughout the training phase to better reflect their naturalistic correspondences. Specifically, our Near-Instance-Wise Attribute Adaptation (NAA) modifies class attribute vectors to obtain multiple attribute basis vectors, generating a subspace that is more relevant to instance-level samples. Additionally, our Vision Attribute Relation Strengthening (VARS) module searches for attribute-related regions within the features, offering additional location information during the training phase. The proposed method is evaluated on four ZSL benchmarks, revealing that it is superior or competitive to the state-of-the-art methods on ZSL and the more challenging Generalized Zero-Shot Learning (GZSL) settings. Extensive experiments corroborate the sustainability of this study as one of the most potential directions for ZSL, i.e., the effectiveness of enhancing the visual-semantic relationships formed during training using a simple model structure. Code is available at: https://github.com/LanchJL/IAB-GZSL.