Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence

Abstract

Product Quantization (PQ) has long been a mainstream for generating an\nexponentially large codebook at very low memory/time cost. Despite its success,\nPQ is still tricky for the decomposition of high-dimensional vector space, and\nthe retraining of model is usually unavoidable when the code length changes. In\nthis work, we propose a deep progressive quantization (DPQ) model, as an\nalternative to PQ, for large scale image retrieval. DPQ learns the quantization\ncodes sequentially and approximates the original feature space progressively.\nTherefore, we can train the quantization codes with different code lengths\nsimultaneously. Specifically, we first utilize the label information for\nguiding the learning of visual features, and then apply several quantization\nblocks to progressively approach the visual features. Each quantization block\nis designed to be a layer of a convolutional neural network, and the whole\nframework can be trained in an end-to-end manner. Experimental results on the\nbenchmark datasets show that our model significantly outperforms the\nstate-of-the-art for image retrieval. Our model is trained once for different\ncode lengths and therefore requires less computation time. Additional ablation\nstudy demonstrates the effect of each component of our proposed model. Our code\nis released at https://github.com/cfm-uestc/DPQ.\n