MultiRes-NetVLAD: Augmenting Place Recognition Training With Low-Resolution Imagery

Visual Place Recognition (VPR) is a crucial component of 6-DoF localization, visual SLAM and structure-from-motion pipelines, tasked to generate an initial list place match hypotheses by matching global descriptors. However, commonly-used CNN-based methods either process multiple image resolutions after training or use single resolution limit multi-scale feature extraction the last convolutional layer during training. In this paper, we augment NetVLAD representation learning with low-resolution pyramid encoding which leads richer representations. The resultant multi-resolution can be conveniently aggregated through VLAD into compact representation, avoiding need for concatenation summation patches in recent approaches. Furthermore, show that underlying learnt tensor combined existing approaches improve their baseline performance. Evaluation on 15 viewpoint-varying viewpoint-consistent benchmarking datasets confirm proposed MultiRes-NetVLAD state-of-the-art Recall@N performance descriptor based retrieval, compared against 11 techniques. Source code publicly available at https://github.com/Ahmedest61/MultiRes-NetVLAD.