CoCALC: A Self-Supervised Visual Place Recognition Approach Combining Appearance and Geometric Information

Visual place recognition (VPR) is considered among the most complicated tasks in SLAM due to multiple challenges of drastic variations both appearance and viewpoint. To address this issue, article presents a self-supervised lightweight VPR approach (namely CoCALC) that fully utilizes geometric information provided by images. The main thing makes CoCALC ultra-lightweight (only 0.27 MB) our use Depthwise Separable Convolution (DSC), simple but effective architecture enables model generate more robust image representation. network trained specifically for can efficiently extract deep convolutional features from salient regions have relatively higher entropy, thereby expanding its applications on resource-limited platforms without GPUs. further eliminate negative consequences high percent false matches, novel band-matrix-based check employed filter out incorrect matching patches, impact different bandwidths recall rate discussed. Results several benchmark datasets confirm proposed yield state-of-the-art performance superior generalization with acceptable efficiency. All relevant codes are at https://github.com/LiKangyuLKY/CoCALC-VPR studies.