Progressive Local Filter Pruning for Image Retrieval Acceleration

Abstract

Most image retrieval works aim at learning discriminative visual features, while little attention is paid to the retrieval efficiency. The speed of feature extraction is key to the real-world system. Therefore, in this paper, we focus on network pruning for image retrieval acceleration. Different from the classification models predicting discrete categories, image retrieval models usually extract continuous features for retrieval, which are more sensitive to network pruning. Such different characteristics of the retrieval and classification models make the traditional pruning method sub-optimal for image retrieval acceleration. Two points are critical for pruning image retrieval models: preserving the local geometry structure of filters and maintaining the model capacity during pruning. In view of the above considerations, we propose a Progressive Local Filter Pruning (PLFP) method. Specifically, we analyze the <italic>local</italic> geometry of filter distribution in every layer and select redundant filters according to one new criterion that the filter can be replaced locally by other similar filters. Furthermore, to preserve the model capacity of the original model, the proposed method <italic>progressively</italic> prune the filter by decreasing the scale of filter weights gradually. We evaluate our method on four scene retrieval datasets, <italic>i.e.</italic>, Oxford5K, Oxford105&#x00A0;K, Paris6K, and Paris106&#x00A0;K, and one person re-identification dataset, <italic>i.e.</italic>, Market-1501. Extensive experiments show that the proposed method (1) preserves the original model capacity while pruning (2) and achieves superior performance to other widely-used pruning methods.