Improving Road Segmentation in Challenging Domains Using Similar Place Priors

Road segmentation in challenging domains, such as night, snow or rain, is a difficult task. Most current approaches boost performance using fine-tuning, domain adaptation, style transfer, by referencing previously acquired imagery. These share one more of three significant limitations: reliance on large amounts annotated training data that can be costly to obtain, both anticipation and from the type environmental conditions expected at inference time, and/or imagery captured previous visit location. In this research, we remove these restrictions improving road based similar places. We use Visual Place Recognition (VPR) find xmlns:xlink="http://www.w3.org/1999/xlink">similar but geographically distinct places, fuse segmentations for query images similar place priors Bayesian approach novel quality metric. Ablation studies show need re-evaluate notions VPR utility demonstrate system achieving state-of-the-art across multiple condition scenarios including night time snow, without requiring any prior access same geographical locations. Furthermore, method network agnostic, improves baseline techniques competitive against methods specialised prediction.