Physical Words for Place Recognition in Dense RGB-D Maps

Appearance-based place recognition systems have been shown to be effective for large-scale mapping but have notable shortcomings. Visual bag-of-words dictionaries require offline training, have tens of thousands of words, and are susceptible to changing environments, either due to lighting or physical changes, between training and deployment. Recent advances allow for online 3D mapping and segmentation using dense RGB-D data. Here we propose the natural extension of previous visual dictionaries to the 3D world through the use of physical words that are used to perform place recognition. The main advantages of this approach is generating and detecting physical words is invariant to aspect and lighting changes, and require less words in our physical dictionary to recognize scenes. We demonstrate this concept on multiple real world datasets under extreme lighting variations and camera trajectories that typical appearance-based approaches have difficulty with.