Attention-Based Accelerated Coordinate Encoding Network for Visual Relocalization

Visual relocalization aims to estimate the 6-degree-of-freedom pose of the camera within a given scene. However, current methods for scene coordinate regression are unable to achieve high localization accuracy and faster training speed simultaneously. Inspired by the trade-off between accuracy and training time, this paper leverages lightweight modules for the accelerated coordinate encoding method to improve localization accuracy. Specifically, the Coordinate Convolution (CoordConv) combines coordinate attention and regular convolution layers, which not only provides position encoding for feature elements but also calculates the importance weight of feature elements based on coordinate attention. Furthermore, a Group Convolutional Block Attention Module (G-CBAM) is added to enhance 3D scene coordinate prediction by leveraging channel and spatial information of the features. The experimental results demonstrate that the proposed method maintains the high efficiency brought by the accelerated coordinate encoding, while improving the position and orientation accuracy by 21% and 25% respectively, compared to the DSAC* on the public cambridge dataset, validating the effectiveness of the proposed approach.