STA-VPR: Spatio-Temporal Alignment for Visual Place Recognition

Recently, the methods based on Convolutional Neural Networks (CNNs) have gained popularity in field of visual place recognition (VPR). In particular, features from middle layers CNNs are more robust to drastic appearance changes than handcrafted and high-layer features. Unfortunately, holistic mid-layer lack robustness large viewpoint changes. Here we split into local features, propose an adaptive dynamic time warping (DTW) algorithm align spatial domain while measuring distance between two images. This realizes viewpoint-invariant condition-invariant recognition. Meanwhile, a matching DTW (LM-DTW) is applied perform image sequence temporal alignment, which achieves further improvements ensures linear complexity. We extensive experiments five representative VPR datasets. The results show that proposed method significantly improves CNN-based methods. Moreover, our outperforms several state-of-the-art maintaining good run-time performance. work provides novel way boost performance CNN without any re-training for VPR. code available at https://github.com/Lu-Feng/STA-VPR.