Research map / Continual learning / Class-incremental
Mitigating Background Shift in Class-Incremental Semantic Segmentation
Gilhan Park, WonJun Moon, SuBeen Lee, Tae-Young Kim, Jae-Pil Heo
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 selective pseudo-labeling (ablation) | Segmenter CISS baseline · Pascal VOC 15-1 overlapped | mIoU (all): 74.2 -> 79.1 | Improved |
| add adaptive feature distillation (ablation) | Segmenter CISS baseline · Pascal VOC 15-1 overlapped | mIoU (all): 74.2 -> 77.0 | Improved |
| add separating objective (LGKD + orthogonal) (ablation) | Segmenter CISS baseline · Pascal VOC 15-1 overlapped | mIoU (new classes 16-20): 49.8 -> 69.6 | Improved |
| add separating objective alone (ablation) | Segmenter CISS baseline · Pascal VOC 15-1 overlapped | mIoU (old classes 1-15): 81.1 -> 80.8 | No clear effect |
| combine all three components (ablation) | Segmenter CISS · Pascal VOC 15-1 overlapped | mIoU (all): 74.2 -> 80.6 | Improved |
| add all components on ADE20k 100-50 (ablation) | Segmenter CISS · ADE20k 100-50 overlapped | mIoU (all): 43.2 -> 45.7 | Improved |
| add AFD or L_Sep on top of SPL on ADE20k (ablation) | SPL model · ADE20k 100-50 overlapped | mIoU (all): 45.2 -> 45.2 (AFD); 45.2 -> 45.5 (L_Sep) | Mixed |
| add orthogonal loss on top of LGKD (ablation) | Background weight transfer model · ADE20k 100-50 overlapped | similarity (C^t, c_0): 90.03 -> 53.86 | Improved |
| add LGKD alone to background weight transfer (ablation) | Background weight transfer model · ADE20k 100-50 overlapped | similarity (C^t, c_0): 91.46 -> 90.03 | No clear effect |
| use MBS instead of INC on multi-step Pascal VOC 15-1 overlapped | Transformer CISS · Pascal VOC 15-1 overlapped | mIoU (all): 75.6 -> 80.6 | Improved |
| use MBS instead of INC on ADE20k | Transformer CISS · ADE20k 100-5 overlapped | mIoU (all): 41.7 -> 42.8 | Improved |
| selective pseudo-labeling ignores ambiguous pixels | Pseudo-labeling · Pascal VOC 15-1 overlapped | ignored label ratio: 37.18 avg ignored ratio | Improved |
Closest papers by meaning
- Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation (ECCV 2024 · similarity 0.93)
- Decomposed Knowledge Distillation for Class-Incremental Semantic Segmentation (NeurIPS 2022 · similarity 0.92)
- Deconfound Semantic Shift and Incompleteness in Incremental Few-shot Semantic Segmentation (AAAI 2025 · similarity 0.88)
- Adaptive Prototype Replay for Class Incremental Semantic Segmentation (AAAI 2025 · similarity 0.88)
- CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation (CVPR 2025 · similarity 0.88)
- Beyond Background Shift: Rethinking Instance Replay in Continual Semantic Segmentation (CVPR 2025 · similarity 0.87)
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.
