2018/01/15 by Sourav Garg, Garg, Sourav, Niko Suenderhauf +3 · 1 citation
Computer Science · Engineering · #Video Surveillance and Tracking Methods #Advanced Image and Video Retrieval Techniques #Automated Road and Building Extraction
paper · pdf · doi:10.48550/arxiv.1801.05078
When a human drives a car along a road for the first time, they later\nrecognize where they are on the return journey typically without needing to\nlook in their rear-view mirror or turn around to look back, despite significant\nviewpoint and appearance change. Such navigation capabilities are typically\nattributed to our semantic visual understanding of the environment [1] beyond\ngeometry to recognizing the types of places we are passing through such as\n"passing a shop on the left" or "moving through a forested area". Humans are in\neffect using place categorization [2] to perform specific place recognition\neven when the viewpoint is 180 degrees reversed. Recent advances in deep neural\nnetworks have enabled high-performance semantic understanding of visual places\nand scenes, opening up the possibility of emulating what humans do. In this\nwork, we develop a novel methodology for using the semantics-aware higher-order\nlayers of deep neural networks for recognizing specific places from within a\nreference database. To further improve the robustness to appearance change, we\ndevelop a descriptor normalization scheme that builds on the success of\nnormalization schemes for pure appearance-based techniques such as SeqSLAM [3].\nUsing two different datasets - one road-based, one pedestrian-based, we\nevaluate the performance of the system in performing place recognition on\nreverse traversals of a route with a limited field of view camera and no\nturn-back-and-look behaviours, and compare to existing state-of-the-art\ntechniques and vanilla off-the-shelf features. The results demonstrate\nsignificant improvements over the existing state of the art, especially for\nextreme perceptual challenges that involve both great viewpoint change and\nenvironmental appearance change. We also provide experimental analyses of the\ncontributions of the various system components.\n