Superparsing: Scalable Nonparametric Image Parsing with Superpixels

摘要: This paper presents a simple and effective nonparametric approach to the problem of image parsing, or labeling regions (in our case, superpixels produced by bottom-up segmentation) with their categories. requires no training, it can easily scale datasets tens thousands images hundreds labels. It works scene-level matching global descriptors, followed superpixel-level local features efficient Markov random field (MRF) optimization for incorporating neighborhood context. Our MRF setup also compute simultaneous into semantic classes (e.g., tree, building, car) geometric (sky, vertical, ground). system outperforms state-of-the-art non-parametric method based on SIFT Flow dataset 2,688 33 In addition, we report per-pixel rates larger 15,150 170 To knowledge, this is first complete evaluation parsing size, establishes new benchmark problem.