Research map / Detection and segmentation / Weakly supervised detection and segmentation
Robots Understanding Contextual Information in Human-Centered Environments Using Weakly Supervised Mask Data Distillation
Daniel Dworakowski, Angus Fung, Goldie Nejat
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 |
|---|---|---|---|
| increase binary search steps from 1 to 6 (2 models, 32 samples) (ablation) | WeSuperMaDD · ICDAR-13 train | F1: +40% | Improved |
| increase binary search steps from 1 to 6 (1 model, 1 sample) (ablation) | WeSuperMaDD · ICDAR-13 train | F1: +210% | Improved |
| add second text recognition model (TPS) to ensemble or raise samples to 32 (one search step) (ablation) | WeSuperMaDD · ICDAR-13 train | F1: +12 to 20 F1 | Improved |
| increase search steps beyond 4 with 32 samples and 2 models (ablation) | WeSuperMaDD · ICDAR-13 train | F1: plateau at approximately 81 at 4 steps | No clear effect |
| WeSuperMaDD pseudo labels instead of GrabCut | PSL generation · ICDAR-13 test | F1: 57.25 -> 80.69 | Improved |
| WeSuperMaDD pseudo labels instead of Pyramid | PSL generation · ICDAR-13 test | F1: 43.12 -> 80.69 | Improved |
| WeSuperMaDD pseudo labels instead of Naive full-box labels | PSL generation · ICDAR-13 test | F1: 67.44 -> 80.69 | Improved |
| weakly supervised WeSuperMaDD instead of semi-supervised SGP | PSL generation · ICDAR-13 test (size-weighted) | F1: 78.87 -> 79.00 | No clear effect |
| train detector with WeSuperMaDD PSLs instead of Naive PSLs (detection) | RetinaNet-style text CNN · ICDAR-13/15/17 and grocery | detection F1: between 0.12 and 1.50 higher | Improved |
| train detector with WeSuperMaDD PSLs instead of Naive PSLs (segmentation) | RetinaNet-style text CNN · ICDAR-13/15/17 and grocery | segmentation F1 after PRP: 41.76 to 71.03 -> 67.10 to 85.24, gain of 42% to 83% | Improved |
| WCNN versus other one-stage detectors | WCNN · ICDAR-13 and ICDAR-15 | detection F1: +0.56 and +3.25 | Improved |
| WCNN versus STELA one-stage detector | WCNN · ICDAR-17 | detection F1: 2.85 lower | Hurt |
Closest papers by meaning
- Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach (AAAI 2020 · similarity 0.86)
- Know Your Attention Maps: Class-specific Token Masking for Weakly Supervised Semantic Segmentation (ICCV 2025 · similarity 0.86)
- Single-Stage Semantic Segmentation From Image Labels (CVPR 2020 · similarity 0.86)
- Weakly Supervised Semantic Segmentation for Driving Scenes (AAAI 2024 · similarity 0.85)
- Context Encoding for Semantic Segmentation (CVPR 2018 · similarity 0.85)
- Causal Intervention for Weakly-Supervised Semantic Segmentation (NeurIPS 2020 · similarity 0.85)
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.
