This paper proposes semantic texton forests as new low-level feature, which is efficient compared with k-means clustering. This paper also presents the bag of semantics textons, which is computed over the whole image for categorization and local rectangular regions for segmentation. Finally, this paper uses an image-level prior for segmentation based on SFT and BOST.
1. Semantic Texton Forests (STF):
a. In training, generate randomized decision trees by the following steps.
i) At root, randomly select a small subset I’ of dataset I.
ii) Spilt into left and right subsets (Il, Ir) by split function f and threshold t, and repeat splitting until leaf node.
iii) Repeat i) and ii) for T times to generate T trees.
b. Feature extraction: a path from root to leaf and a class distribution at leaf.
2. Bags of Semantic Textons (BOST):
a. A prior estimate in a given region (the region could be the whole image).
b. Semantic texton histogram: counts of each visited node of every pixels in the region.
c. Region prior: average class distribution of each visited leaf node.
3. Image-level Prior (ILP):
a. Emphasize the likely categories and discourage unlikely categories.
b. Multiply the distributions by parameter α to soften the prior.
沒有留言:
張貼留言