Channel-Wise Part-Based Weighting Aggregation for Image Retrieval

The part-based weighting aggregation (PWA) method is a simple and straightforward way for creating powerful image representations in image retrieval. In this paper we analyze the burstiness among the responses of the aggregated channels in the PWA method, and propose a channel-wise weights utilizing the sum-pooling aggregation of the deep features' responses to regulate the burstiness among the feature maps. Firstly, we apply the channel weights to the aggregation stage of PWA. The regulation of the burstiness among the aggregated channels boosts the PWA performance. Secondly, we use the channel weights to the part selection stage, namely, selecting the part detectors based on the channel weighted feature maps to obtain more discriminate parts to further improve the PWA performance. Comprehensive experiments on four standard datasets have demonstrated the effects of the proposed method.