Research map / Detection and segmentation / Weakly supervised detection and segmentation
Learning from the Web: Language Drives Weakly-Supervised Incremental Learning for Semantic Segmentation
Chang Liu, Giulia Rizzoli, Pietro Zanuttigh, Fu Li, Yi Niu
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 |
|---|---|---|---|
| add Fourier-amplitude domain filtering of web images (ablation) | WILSON with web data · PASCAL-VOC 10-10 overlapped | mIoU (all): 62.1 -> 63.1 | Improved |
| add caption labeling for multi-label image-level supervision (ablation) | WILSON with web data · PASCAL-VOC 10-10 overlapped | mIoU (all): 62.1 -> 63.6 | Improved |
| combine domain filtering and caption labeling (ablation) | WILSON with web data · PASCAL-VOC 10-10 overlapped | mIoU (all): 62.1 -> 65.0 | Improved |
| query rehearsal images by caption instead of class name (ablation) | web rehearsal · PASCAL-VOC 10-10 overlapped | mIoU (all): 63.6 -> 64.9 | Improved |
| add caption-based filtering of rehearsal images (ablation) | web rehearsal · PASCAL-VOC 10-10 overlapped | mIoU (all): 63.6 -> 63.5 | No clear effect |
| combine caption query and caption filtering (ablation) | web rehearsal · PASCAL-VOC 15-5 overlapped | mIoU (all): 69.8 -> 73.4 | Improved |
| Fourier amplitude discriminator instead of pixel-domain discriminator (ablation) | web image selection · PASCAL-VOC 10-10 overlapped | mIoU (all): 61.9 -> 63.1 | Improved |
| use 500 instead of 100 web images for rehearsal or new-class training (ablation) | web-learning pipeline · PASCAL-VOC 10-10 overlapped | mIoU (all): 65.7 (M=100) vs 65.3 (M=500) | Mixed |
| replace OpenFlamingo with BLIP captioner (ablation) | Ours · PASCAL-VOC 10-10 and 15-5 overlapped | mIoU (all): 65.7 -> 65.9 (10-10); 73.3 -> 73.0 (15-5) | No clear effect |
| vary caption-filter threshold T from 0.5 to 0.7 (ablation) | caption filtering · PASCAL-VOC 10-10 and 15-5 overlapped | mIoU (all): T=0.6 gives 65.7 and 73.3 | No clear effect |
| use web rehearsal data for multi-step single-class incremental learning | WILSON · PASCAL-VOC 10-1 overlapped | mIoU (all): 0.1 -> 54.4 | Improved |
Closest papers by meaning
- Foundation Model Drives Weakly Incremental Learning for Semantic Segmentation (CVPR 2023 · similarity 0.92)
- Tendency-driven Mutual Exclusivity for Weakly Supervised Incremental Semantic Segmentation (ECCV 2024 · similarity 0.92)
- Incremental Learning in Semantic Segmentation From Image Labels (CVPR 2022 · similarity 0.91)
- Modeling the Background for Incremental and Weakly-Supervised Semantic Segmentation (TPAMI 2021 · similarity 0.90)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.88)
- Deconfound Semantic Shift and Incompleteness in Incremental Few-shot Semantic Segmentation (AAAI 2025 · similarity 0.88)
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.
