MS-MixVPR: Multi-scale Feature Mixing Approach for Long-Term Place Recognition

Visual place recognition (VPR) is a crucial task in robotics and autonomous systems, enabling robots to localize themselves in complex and dynamic environments. Due to significant differences in appearance that arise from changes in environmental factors like season, weather, and lighting (day or night), VPR is particularly challenging in outdoor settings. This paper presents a novel method to address this challenge called MS-MixVPR, which is proposed based on an existing work, MixVPR. The proposed MS-MixVPR extracts global features from different layers of pre-trained CNN backbones using MixVPR’s Feature Mixer blocks. These visual cues are combined further to create a compact, holistic representation that is highly robust to changes in environmental conditions. We evaluate the proposed MS-MixVPR on four challenging real-world benchmark datasets, including Nordland, SPEDTest, MSLS, and Pittsburgh30k. The experimental results show that our MS-MixVPR outperforms several current state-of-the-art methods while maintaining low computational time. Consequently, our approach is suitable for real-world applications that are often resource-constrained.