Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly-Supervised Semantic Segmentation Using Motion Cues
Pavel Tokmakov, Karteek Alahari, Cordelia Schmid
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 soft motion constraints instead of hard motion labels (ablation) | M-CNN* · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 29.9 -> 33.6 | Improved |
| add fine-tuning on selected reliable shots (ablation) | M-CNN (large FOV) · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 33.6 -> 41.2 | Improved |
| use large field of view (224x224) instead of small (128x128) (ablation) | M-CNN* · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 26.7 -> 33.6 | Improved |
| use large field of view instead of small (ablation) | EM-Adapt · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 23.2 -> 17.9 | Hurt |
| replace EM-Adapt size heuristics with motion-based label prediction | M-CNN vs EM-Adapt · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 17.9 -> 41.2 | Improved |
| add weakly-annotated VOC images to YouTube-Objects videos (ablation) | M-CNN · VOC 2012 val (10 classes) | mean IoU: 41.2 -> 47.2 | Improved |
| use network prediction only for VOC images instead of EM-Adapt (ablation) | M-CNN (VOC+YouTube-Objects) · VOC 2012 val (10 classes) | mean IoU: 47.2 -> 51.0 | Improved |
| add VOC images to ImageNet-VID videos (ablation) | M-CNN · VOC 2012 val (10 classes) | mean IoU: 39.0 -> 43.7 | Improved |
| train with ImageNet-VID instead of YouTube-Objects | M-CNN · VOC 2012 val (10 classes) | mean IoU: 41.2 -> 39.0 | No clear effect |
| vary number of graph-cut iterations from 1 to 5 (ablation) | M-CNN* · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 33.6, 33.1, 33.5, 33.6, 33.9 | No clear effect |
| vary alpha from 0.5 to 10.0 (ablation) | M-CNN* · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 24.7 (0.5), 33.8 (1.0), 34.3 (3.0), 23.3 (10.0) | Mixed |
| vary alpha in fine-tuning from 0.5 to 10.0 (ablation) | M-CNN · VOC 2012 val (10 classes, trained on YouTube-Objects) | mean IoU: 41.4 (0.5), 41.9 (1.0), 42.6 (3.0), 42.2 (10.0) | No clear effect |
Closest papers by meaning
- Combining Bottom-Up, Top-Down, and Smoothness Cues for Weakly Supervised Image Segmentation (CVPR 2017 · similarity 0.89)
- Bringing Background Into the Foreground: Making All Classes Equal in Weakly-Supervised Video Semantic Segmentation (ICCV 2017 · similarity 0.89)
- Seed, Expand and Constrain: Three Principles for Weakly-Supervised Image Segmentation (ECCV 2016 · similarity 0.89)
- Built-in Foreground/Background Prior for Weakly-Supervised Semantic Segmentation (ECCV 2016 · similarity 0.89)
- Webly Supervised Semantic Segmentation (CVPR 2017 · similarity 0.89)
- Revisiting Dilated Convolution: A Simple Approach for Weakly- and Semi-Supervised Semantic Segmentation (CVPR 2018 · 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.
