Research map / Continual learning / Class-incremental
CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoMBO, a Mask2Former-based class-incremental segmentation method that branches queries with class-specific low-rank adapters (QCR) and adds half-distillation-half-learning and importance-weighted query distillation to reduce old/new class conflict on ADE20K.
Problem
Class incremental segmentation (panoptic and semantic) suffers from a conflict between retaining old classes (catastrophic forgetting) and learning new classes (plasticity), with existing methods trading one off against the other.
Approach
Decouple learning and distillation at the query level: class-specific low-rank adapters refine queries for new classes while original queries are preserved for distillation; the classification loss on unmatched queries is replaced by a shared KL distillation over old and new logits (HDHL), and query distillation is weighted by a per-query importance score (IKD).
Key results
- Achieves 35.6% PQ on ADE20K CIPS 100-10, the best overall score among compared exemplar-free methods, and 41.1% mIoU on CISS 100-10.
- Ablation on CIPS 100-10: full CoMBO reaches 35.6 overall PQ versus 32.2 for the baseline with pseudo-labeling (+3.4).
- QCR adds about 8.2K parameters per class at r=16, roughly 2% of Mask2Former's 44.9M parameters with ResNet-50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HDHL to vanilla losses with pseudo-labeling (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 32.2 -> 34.3 | Helped |
| Add IKD to HDHL (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 34.3 -> 34.9 | Helped |
| Add QCR to HDHL and IKD (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 34.9 -> 35.6 | Helped |
| Use HDHL instead of vanilla Lcls on all classification embeddings (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 32.2 -> 34.3 | Helped |
| Use auxiliary distillation on embeddings (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 32.2 -> 32.6 | Helped |
| Use distillation only for unmatched embeddings (ablation) | CoMBO on ADE20K CIPS 100-10 · ADE20K CIPS 100-10 (6 steps) | PQ: 32.2 -> 32.2 | No clear effect |
Surprising findings
- On CIPS 100-5, CoMBO underperforms ECLIPSE overall, which the authors attribute to ECLIPSE's stronger memorization from freezing parameters.
Limitations
- Evaluated only on ADE20K.
- Underperforms ECLIPSE on overall PQ in the 100-5 scenario.
Experimental setup
Datasets: ADE20K
Benchmarks: ADE20K
Metrics: PQ mIoU
Baselines: FT MiB PLOP CoMFormer ECLIPSE BalConpas CoMasTRe
Builds on: Mask2Former CoMFormer PLOP BalConpas
Closest papers by meaning
- Adaptive Prototype Replay for Class Incremental Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Rethinking Query-based Transformer for Continual Image Segmentation (CVPR 2025 · similarity 0.86)
- Continual Segmentation with Disentangled Objectness Learning and Class Recognition (CVPR 2024 · similarity 0.85)
- $\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation (NeurIPS 2026 · similarity 0.85)
- Relaxed Class-consensus Consistency for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.85)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · 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.