Learning Deep Representations of Fine-grained Visual Descriptions

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.

PDF Abstract CVPR 2016 PDF CVPR 2016 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Few-Shot Image Classification CUB-200-2011 - 0-Shot Word CNN-RNN (DS-SJE Embedding) Top-1 Accuracy 56.8% # 1
AP50 48.7 # 1
Few-Shot Image Classification CUB 200 50-way (0-shot) DS-SJE Reed et al. (2016) Accuracy 50.4 # 3
Few-Shot Image Classification CUB 200 50-way (0-shot) DA-SJE Reed et al. (2016) Accuracy 50.9 # 2
Few-Shot Image Classification Flowers-102 - 0-Shot Word CNN-RNN (DS-SJE Embedding) AP50 59.6 # 1
Accuracy 65.6% # 1

Methods


No methods listed for this paper. Add relevant methods here