Evaluation of Global Descriptor Methods for Appearance-Based Visual Place Recognition

Visual place recognition (VPR) is considered among the most challenging problems due to the extreme variations in appearance and viewpoint. Essentially, appearance-based VPR can be considered as an image retrieval task, thus the key is to accurately and efficiently describe the images. Recently, global descriptor methods have attracted substantial attention from the VPR community, which has contributed to numerous important outcomes. Despite the growing number of global descriptors presented, little attention has been paid to the comparison and evaluation of these methods and so it remains difficult for researchers to disentangle the factors that led to better performance. This study provided comprehensive insight into global descriptors from a practical application perspective. We present a systematic evaluation that integrates 15 commonly used global descriptors, 6 benchmark datasets, and 5 evaluation metrics, and subsequently extended this evaluation to discuss the key factors impacting the matching performance and computational efficiency. We also report practical suggestions for constructing promising CNN descriptors, based on the experimental conclusions. Our analysis reveals both advantages and limitations of three different types of global descriptors, including handcrafted features-based ones, off-the-shelf CNN-based ones, and customized CNN-based ones. Finally, we evaluate the practicality of reported global descriptors to mediate the trade-offs between matching performance and computational efficiency.