Research map / Detection and segmentation / Open-set and few-shot detection

Deconfound Semantic Shift and Incompleteness in Incremental Few-shot Semantic Segmentation

AAAI 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Few-shot segmentation · 2 citations (OpenAlex)

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 a causal-inference framework for incremental few-shot semantic segmentation with a Causal Intervention Module (CIM) against semantic shift and a Prototype Refinement Module (PRM) that fuses old/new prototypes via episodic knowledge to counter incompleteness.

Problem

Incremental few-shot semantic segmentation (IFSS) suffers from semantic shift (background and object classes swapping across steps) and semantic incompleteness (few new-class samples lacking representative attributes), causing forgetting and poor new-class learning.

Approach

Models IFSS with a structural causal model and deconfounds it via front-door-style intervention on old-class prototypes (CIM, prototype-attention reweighting with no training) and a meta-learned encoder-decoder over episodes that fuses old and new prototypes to complete new-class prototypes (PRM).

Key results

• On PASCAL-VOC 2012 15-5 (T=2), the method reaches 76.30 mIoU on old classes and 50.67 on new classes (full), with 34.20 (5-shot) and 21.37 (1-shot).
• On ADE20K 100-50, the method reaches 10.41 (5-shot) and 6.84 (1-shot) new-class mIoU, outperforming PLOP and MiB.
• Adding CIM to EHNet improves old-class mIoU by 14.27 points on VOC 15-5; adding PRM to PLOP raises new-class 1-shot mIoU on ADE20K 50-50 by about 46 points relative to baseline.

Measured effects reported in the paper

Surprising findings

• The EM replay baseline, which stores old-class data, achieves higher old-class mIoU than the proposed method in several settings.
• PRM embedded into PLOP at 1-shot on ADE20K 100-10 gives the largest new-class gains, while PLOP+PRM reduces some old-class metrics.

Limitations

• Performance on old classes and new classes remains low in long-term settings (e.g., 15-1 T=6 and ADE20K 100-10 1-shot new-class mIoU of 4-7).
• Still fails in some cases, e.g., forgetting old class 'horse' under small imbalanced new-class data, as noted in qualitative results.

Experimental setup

Datasets: PASCAL-VOC 2012 ADE20K

Benchmarks: PASCAL-VOC 2012 ADE20K

Metrics: mIoU

Baselines: FT GIFS PIFS EHNet ILT MiB EM PLOP

Builds on: MiB PLOP PIFS EHNet masked average pooling

Closest papers by meaning

• Few-Shot Incremental Learning via Foreground Aggregation and Knowledge Transfer for Audio-Visual Semantic Segmentation (AAAI 2025 · similarity 0.87)
• Adaptive Prototype Replay for Class Incremental Semantic Segmentation (AAAI 2025 · similarity 0.87)
• Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.86)
• Enhancing Generalized Few-Shot Semantic Segmentation via Effective Knowledge Transfer (AAAI 2025 · similarity 0.86)
• APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation (CVPR 2024 · similarity 0.86)
• Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.86)