Semantic Segmentation Refinement by Monte Carlo Region Growing of High Confidence Detections
Document Type
Conference Proceeding
Publication Date
6-2-2019
Publisher
Springer
Source Publication
Lecture Notes in Computer Science
Source ISSN
0302-9743
Abstract
The semantic segmentation produced by most state-of-the-art methods does not show satisfactory adherence to object boundaries. Methods such as fully-connected conditional random fields (CRFs) can significantly refine segmentation predictions. However, they rely on supervised parameter optimization that depends upon specific datasets and predictor modules. We propose an unsupervised method for semantic segmentation refinement that takes as input the confidence scores generated by a segmentation network and re-labels pixels with low confidence levels. More specifically, a region growing mechanism aggregates these pixels to neighboring areas with high confidence scores and similar appearance. To minimize the impact of high-confidence prediction errors, our algorithm performs multiple growing steps by Monte Carlo sampling initial seeds in high-confidence regions. Our method provides both running time and segmentation improvements comparable to state-of-the-art refinement approaches for semantic segmentation, as demonstrated by evaluations on multiple publicly available benchmark datasets.
Recommended Citation
Ambrozio Dias, Philipe and Medeiros, Henry, "Semantic Segmentation Refinement by Monte Carlo Region Growing of High Confidence Detections" (2019). Electrical and Computer Engineering Faculty Research and Publications. 698.
https://epublications.marquette.edu/electric_fac/698
Comments
Lecture Notes in Computer Science, Vol. 11362 (2018). DOI.