Deep Fusion of Multi-attentive Local and Global Features with Higher Efficiency for Image Retrieval

Baorong Shi

Keywords: Image retrieval, Homography learning, Attention, Intermediate supervision

Abstract: Image retrieval is to search images similar to the given query image by extracting features. Previously, methods that firstly search by global features then re-rank images using local feature matching were proposed, which has an excellent performance on many datasets. However, their drawbacks are also obvious. For example, the local feature matching consumes time and space greatly, the re-ranking process weakens the influence of global features, and the local feature learning is not accurate enough and semantic enough because of the trivial design. In this work, we proposed a Unifying Global and Attention-based Local Features Retrieval method (referred to as UGALR), which is an end-to-end and single-stage pipeline. Particularly, UGALR benefits from two aspects: 1) it accelerates extraction speed and reduces memory consumption by removing the re-ranking process and learning local feature matching with convolutional neural networks instead of RANSAC algorithm; 2) it learns more accurate and semantic local information through combining spatial and channel attention with the aid of intermediate supervision. Experiments on Revisited Oxford and Paris datasets validate the effectiveness of our approach, and we achieved state-of-the-art performance compared to other popular methods. The codes will be available soon.