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

On the Relaxation of Conditional Independence Assumption for Image Segmentation

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Interactive and referring segmentation · 0 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.

Replaces RankSEG's conditional independence assumption with a spatially localized Gaussian-kernel covariance model, and uses reciprocal moment approximation plus a fixed-point search to optimize expected Dice/IoU in O(d log d) at inference time.

Problem

RankSEG optimizes Dice/IoU at inference under conditional independence of pixel labels, which ignores label correlations and hurts performance in low-contrast or ambiguous regions; full dependence modeling costs O(d^3).

Approach

Model label covariance as spatially localized (Gaussian-kernel decay with distance), approximate the reciprocal moment by the inverse of the first moment so the needed conditional means become a convolution computed via FFT, and find the optimal predicted volume by a fixed-point iteration over a single global ranking instead of exhaustive re-ranking.

Key results

• On KiTS with DeepLabV3+, the method reaches 64.07 Dice vs 61.16 for Argmax-prob and 63.56 for CIA-RankSEG.
• On ADE20K with UPerNet, mDice improves from 63.98 (Argmax-prob) and 64.92 (CIA-RankSEG) to 65.19.
• The fixed-point iteration terminates within 2.5 steps for over 99.8% of test samples across five benchmarks.

Surprising findings

• CRF post-processing consistently underperforms simple Argmax-prob on the multi-class benchmarks.
• On LiTS the fixed-point solution essentially matches brute-force global optimum (49.88 Dice vs 49.88).

Limitations

• SLD uses only spatial distance and ignores appearance cues such as image intensity or prediction similarity.
• The approximation relies on a single bandwidth hyperparameter theta and on assumptions such as mu_j/d = Theta(1) for the error bound.
• Convergence to the optimum is supported empirically rather than with a general guarantee; the single-step condition is hard to verify on real data.

Experimental setup

Datasets: LiTS KiTS DeepGlobe Land ADE20K Cityscapes

Benchmarks: LiTS KiTS ADE20K Cityscapes DeepGlobe

Metrics: Dice IoU mDice mIoU

Baselines: Argmax-prob CIA-RankSEG CRF

Builds on: RankSEG reciprocal moment approximation conditional random fields

Closest papers by meaning

• RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.90)
• Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.88)
• RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
• Integrating Low-Level Visual Cues for Enhanced Unsupervised Semantic Segmentation (AAAI 2025 · similarity 0.85)
• Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.85)
• Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.85)