GND-APR: Absolute pose regressor with graph neural diffusion for self-driving

Abstract

Visual relocalization consists of estimating the camera translation and rotation in known scenarios. However, the lack of robust features occurs in pose regressors under self-driving environments containing perturbations (such as weather changes, seasons, illumination, and dynamic objects), leading to inaccurate poses. In this paper, we propose GND-APR to cope with the aforementioned issue via Graph dynamic attention Neural Diffusion and shared memory unit. Specifically, we propose graph dynamic attention neural diffusion for feature map and vector interaction, which graph dynamic attention facilitates better information interaction between feature correlations and enhances feature representation. Meanwhile, we enhance temporal feature fusion through a shared memory unit, it allows the network to record essential and robust features as an implicit defense against perturbations. Our model decreases the mean translation and rotation errors by 54% and 15%, respectively, on the 4Seasons dataset. Experiments on two challenging self-driving datasets demonstrate the superiority of our approach over other state-of-the-art methods.