Seed, Expand and Constrain: Three Principles for Weakly-Supervised Image Segmentation
About
We introduce a new loss function for the weakly-supervised training of semantic image segmentation models based on three guiding principles: to seed with weak localization cues, to expand objects based on the information about which classes can occur in an image, and to constrain the segmentations to coincide with object boundaries. We show experimentally that training a deep convolutional neural network using the proposed loss function leads to substantially better segmentations than previous state-of-the-art methods on the challenging PASCAL VOC 2012 dataset. We furthermore give insight into the working mechanism of our method by a detailed experimental study that illustrates how the segmentation quality is affected by each term of the proposed loss function as well as their combinations.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Semantic segmentation | PASCAL VOC 2012 (val) | Mean IoU50.7 | 2040 | |
| Semantic segmentation | PASCAL VOC 2012 (test) | mIoU51.7 | 1342 | |
| Semantic segmentation | CamVid (test) | mIoU2.5 | 411 | |
| Semantic segmentation | COCO 2014 (val) | mIoU22.4 | 251 | |
| Weakly supervised semantic segmentation | PASCAL VOC 2012 (test) | mIoU51.7 | 158 | |
| Weakly supervised semantic segmentation | PASCAL VOC 2012 (val) | mIoU50.7 | 154 | |
| Semantic segmentation | COCO (val) | mIoU22.4 | 135 | |
| Semantic segmentation | COCO Object (val) | mIoU0.224 | 77 | |
| Weakly supervised semantic segmentation | MS-COCO 2014 (val) | mIoU22.4 | 27 | |
| Weakly supervised semantic segmentation | VOC 2012 (val) | mIoU50.7 | 19 |