Abstract
Fine-grained generalized zero-shot learning (GZSL) tasks require exploration of relevance between local visual features and attributes to discover fine distinctive information for satisfactory performance. Embedding learning and feature synthesizing are two of the popular categories of GZSL methods. However, these methods do not explore fine discriminative information as they ignore either the local features or direct guidance from the attributes. Consequently, they do not perform well. We propose a novel embedding learning network with a two-step dense attention mechanism, which uses direct attribute supervision to explore fine distinctive local visual features for fine-grained GZSL tasks. We further incorporate a feature synthesizing network, which uses the attribute-weighted visual features from the embedding learning network. Both networks are mutually trained in an end-to-end fashion to exploit mutually beneficial information. Consequently, the proposed method can test both scenarios: when only the images of unseen classes are available (using the feature synthesizing network) or when both images and semantic descriptors of the unseen classes are available (via the embedding learning network). Moreover, to reduce bias towards the source domain during testing, we compute source-target class similarity based on mutual information and transfer-learn the target classes. We demonstrate that our proposed method outperforms contemporary methods on benchmark datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2101.02141
