Multi-view adversarial domain adaptation for visual place recognition

Visual place recognition is crucial in robot localization and autonomous driving. Most existing methods focus on inter-domain appearance changes, but struggle with intra-domain shifts and viewpoint discrepancy. We propose a multi-view adversarial adaptation for visual place recognition. Our approach uses random augmentation strategy for diversification guided by explicit prior knowledge about the shifts on a specific source. Following this, a fine-grained domain adaptation framework is presented to minimize the intra-domain appearance discrepancy through exploiting these diversified source images. Furthermore, we design a multi-view collaborative learning network by utilizing the correlation of multi-view features from distinct source images to deal with viewpoint discrepancy. This alignment network fully explores the essential geometric information across source images in a mutual learning manner. Subsequently, the semantic reinforced place representation method is developed to embed dynamic robustness into learned global descriptors with multi-scale attention. Finally, we fuse shared beneficial features derived from multiple view collaborators, results in optimal exploration and utilization of domain-invariant features. Extensive experiments on multiple benchmarks demonstrate the effectiveness of our technique. We achieve a high average Recall @1 score of 93. 3% on Ford Multi-AV, 91. 7% on Oxford RobotCar, and 91. 3% on NCLT.