Visual place recognition is a challenging problem due to the vast range of ways in which the appearance of real-world places varies, the major problems come from:
- Illumination changes, scenes under light conditions may look totally unfamiliar.
- Appearance changes, whether or season caused appearance change, such as trees under summer and winter will show a different outlook.
- Viewpoints changes, the same accurate pose may never occur twice, the extract features need to be viewpoints invariant, however, capture the description for a specific scene at the same time.
In our visual place recognition method, we proposed a multi-channel unsupervised feature extraction method, which use the generative adversarial feature to obtain the joint features under 2D top-view maps and 3D local maps, where each map has their advantage in scene representation. Then, we use such joint feature to achieve image retrieval, the experiments show the matching accuracy could still be guaranteed even under huge viewpoints difference. the relative video result could be checked below.