Learning Deep Representations of Fine-grained Visual Descriptions
This work addresses the problem of fine-grained visual recognition for researchers and practitioners by providing a more flexible and compact language-based approach, though it is incremental as it builds on existing joint embedding frameworks.
The paper tackled the limitations of attribute-based methods in zero-shot visual recognition by proposing neural language models trained from scratch on raw text, achieving strong performance in zero-shot text-based image retrieval and significantly outperforming attribute-based state-of-the-art on the Caltech UCSD Birds 200-2011 dataset.
State-of-the-art methods for zero-shot visual recognition formulate learning as a joint embedding problem of images and side information. In these formulations the current best complement to visual features are attributes: manually encoded vectors describing shared characteristics among categories. Despite good performance, attributes have limitations: (1) finer-grained recognition requires commensurately more attributes, and (2) attributes do not provide a natural language interface. We propose to overcome these limitations by training neural language models from scratch; i.e. without pre-training and only consuming words and characters. Our proposed models train end-to-end to align with the fine-grained and category-specific content of images. Natural language provides a flexible and compact way of encoding only the salient visual aspects for distinguishing categories. By training on raw text, our model can do inference on raw text as well, providing humans a familiar mode both for annotation and retrieval. Our model achieves strong performance on zero-shot text-based image retrieval and significantly outperforms the attribute-based state-of-the-art for zero-shot classification on the Caltech UCSD Birds 200-2011 dataset.