Research map / Detection and segmentation / Weakly supervised detection and segmentation
Revisiting Dilated Convolution: A Simple Approach for Weakly- and Semi-Supervised Semantic Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDC rather than Hong et al.'s image-level-label method | Weakly supervised semantic segmentation · PASCAL VOC 2012 validation | mIoU: 58.1 -> 60.4 | Helped |
| Use MDC rather than Hong et al.'s image-level-label method | Weakly supervised semantic segmentation · PASCAL VOC 2012 test | mIoU: 58.7 -> 60.8 | Helped |
| Use MDC rather than WSSL with image-level labels | Semi-supervised semantic segmentation · PASCAL VOC 2012 validation | mIoU: 64.6 -> 65.7 | Helped |
| Use MDC rather than WSSL with image-level labels | Semi-supervised semantic segmentation · PASCAL VOC 2012 test | mIoU: 66.2 -> 67.6 | Helped |
| Increase the single-block dilation rate from d=1 to d=9 (ablation) | MDC localization-derived masks · PASCAL VOC 2012 validation | mIoU: 50.3 -> 54.4 | Helped |
| Fuse d=1 with the average of d=3, 6, and 9 localization maps (ablation) | MDC localization-derived masks · PASCAL VOC 2012 validation | mIoU: 54.4 -> 57.1 | Helped |
| Apply CRF post-processing to fused segmentation (ablation) | MDC weakly supervised segmentation · PASCAL VOC 2012 validation | mIoU: 57.1 -> 60.4 | Helped |
| Average all localization maps, including d=1, instead of preserving the d=1 map and averaging only dilated maps (ablation) | MDC localization fusion · PASCAL VOC 2012 validation | mIoU: drops almost 1% | Hurt |
| Use four convolution blocks with the same dilation rate d=1 instead of one block (ablation) | MDC localization network · PASCAL VOC 2012 validation | mIoU | No clear effect |
| Reduce strongly annotated images from 1.4K to 500 (ablation) | MDC semi-supervised segmentation · PASCAL VOC 2012 validation | mIoU: 62.7 -> 61.8 | Hurt |
| Increase strongly annotated images from 1.4K to 2.9K (ablation) | MDC semi-supervised segmentation with CRF · PASCAL VOC 2012 test | mIoU: 67.6 -> 68.5 | Helped |
Closest papers by meaning
- Combining Bottom-Up, Top-Down, and Smoothness Cues for Weakly Supervised Image Segmentation (CVPR 2017 · similarity 0.92)
- Weakly-Supervised Semantic Segmentation by Iteratively Mining Common Object Features (CVPR 2018 · similarity 0.91)
- Universal Weakly Supervised Segmentation by Pixel-to-Segment Contrastive Learning (ICLR 2021 · similarity 0.90)
- On the Importance of Label Quality for Semantic Segmentation (CVPR 2018 · similarity 0.89)
- Learning Multi-Modal Class-Specific Tokens for Weakly Supervised Dense Object Localization (CVPR 2023 · similarity 0.89)
- Frequency-Adaptive Dilated Convolution for Semantic Segmentation (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.