Figure 3
From: Scene Segmentation with Low-Dimensional Semantic Representations and Conditional Random Fields
Sample segmentation results for the MSRC-21 dataset. The first 8 rows show some representative cases; the last 5 cases are where the methods worked less well. Column (a) shows the original input images. Column (b) shows results from the first stage LRC, mapped to pixel level using the nearest-patch-center method. Columns (c) and (d) show results from the two-stage method LRC/LRC, mapped to pixel level using, respectively, the nearest-patch and MRF smoothing methods. Columns (e) and (f) show results from the two-stage method LRC/CRF, mapped to pixel level using, respectively, the nearest-patch and oversegmentation methods. Column (g) shows the hand-labeled ground truth.