Sequential Descriptors for Visual Place Recognition: Combining EMLA Training With TFA-Net Aggregation

In this paper, we present a novel algorithm aimed at enhancing Visual Place Recognition (VPR) by addressing the inherent limitations of existing sequence-based methods. Our primary contributions encompass three key areas: firstly, we propose a novel training approach, the Enhanced Metric Learning Approach (EMLA), for extracting more robust global descriptors; secondly, we design an advanced sequence aggregation method, the Temporal Frame Aggregation Network (TFA-Net), that effectively integrates information from image sequences; thirdly, we develop a similarity-based descriptor sorting mechanism, Cascade Descriptor Matching with Similarity-Based Descriptor Sorting (CDM-SBDS), to improve matching accuracy and efficiency. We conducted comprehensive evaluations of our proposed methods on four diverse datasets: Oxford, Norland-SF, Amman, and Austin. The experimental results demonstrate that our methods consistently perform well across various scenarios, particularly excel in large-scale image retrieval tasks. Additionally, we provide insights into the impact of sequence length on performance and discuss the computational efficiency of our approach. These results also highlight potential areas for further improvement in handling complex datasets, providing valuable directions for future research.