Research map / Detection and segmentation / CNN features
RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation
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 RankSEG-RMA, which uses a reciprocal moment approximation to cut RankDice/RankIoU decoding to O(d) and adds an RMA-based pixel score for non-overlapping multiclass segmentation.
Problem
Consistent segmentation rules that directly optimize image-level Dice/IoU (RankSEG) are computationally expensive (RankDice O(d log d) with large constant, RankIoU O(d^2)) and only apply to overlapping multiclass settings, whereas standard benchmarks are non-overlapping.
Approach
Approximates the reciprocal moment of a Poisson-binomial volume variable by plugging in its mean (with provable bounds), turning the optimal-volume search into a linear-time computation, and resolves overlaps with a Dice/IoU-based RMA marginal-gain score instead of raw probabilities.
Key results
- RankDice-RMA improves mIoU^I over Argmax-prob on PASCAL VOC, Cityscapes and ADE20K across all six models, e.g. ADE20K UPerNet 56.94 to 57.67 mIoU^I.
- On LiTS, RankDice-RMA runs in 0.34 s versus 16.33 s for RankDice-BA (about 48x speedup) with near-identical Dice^I.
- Ten-seed t-tests on UPerNet give p-values below 1e-6 for RankDice-RMA over Argmax-prob on VOC, Cityscapes and ADE20K.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Argmax-prob with RankDice-RMA | UPerNet (ConvNeXt) · PASCAL VOC | mIoU^I: 87.82 -> 88.25 | Helped |
| Replace Argmax-prob with RankDice-RMA | UPerNet (ConvNeXt) · Cityscapes | mIoU^I: 75.66 -> 76.17 | Helped |
| Replace Argmax-prob with RankDice-RMA | UPerNet (ConvNeXt) · ADE20K | mIoU^I: 56.94 -> 57.67 | Helped |
| Replace RankDice-BA with RankDice-RMA | DeepLabV3+ (ResNet50) · LiTS | IoU^I: 40.09 -> 40.09 | No clear effect |
| Replace RankDice-BA with RankDice-RMA | DeepLabV3+ (ResNet50) · KiTS | IoU^I: 63.56 -> 63.56 | No clear effect |
| Use RMA-scores instead of probability scores to resolve overlaps (ablation) | RankDice-RMA multiclass · PASCAL VOC | mIoU^I: 87.83 -> 88.25 | Helped |
| Use RMA-scores instead of weighted probability scores to resolve overlaps (ablation) | RankDice-RMA multiclass · Cityscapes | mIoU^I: 75.89 -> 76.17 | Helped |
| Use RMA-scores instead of weighted probability scores to resolve overlaps (ablation) | RankDice-RMA multiclass · ADE20K | mIoU^I: 56.75 -> 57.67 | Helped |
| Use weighted probability scores instead of probability scores to resolve overlaps (ablation) | RankDice-RMA multiclass · ADE20K | mIoU^I: 56.96 -> 56.75 | Hurt |
| Use the upper RMA bound instead of the lower-bound approximation (ablation) | RankSEG-RMA · Reported experiments | segmentation performance | No clear effect |
| Replace RankDice-BA with RankDice-RMA | RankSEG prediction on LiTS · LiTS | prediction time: 48x speedup | Helped |
| Replace RankDice-BA with RankDice-RMA | RankSEG prediction on KiTS · KiTS | prediction time: 38x speedup | Helped |
Surprising findings
- RankDice-RMA also improves IoU even though it is derived from the Dice Bayes rule.
- The RMA upper and lower bounds yield the same performance in practice.
Limitations
- Overlap resolution predicts each pixel independently (greedy), which may not be globally optimal.
- Relies on the conditional independence assumption of the Bayes rule.
- RankSEG-RMA is only as good as the probability estimates from the underlying model.
Experimental setup
Datasets: PASCAL VOC Cityscapes ADE20K LiTS KiTS
Benchmarks: PASCAL VOC Cityscapes ADE20K LiTS KiTS
Metrics: mIoU^I mIoU^C mDice^I mDice^C IoU^I Dice^I mIoU^I_5 mIoU^I_10 runtime
Baselines: Argmax-prob RankDice-BA Prob-scores WProb-scores
Builds on: RankSEG RankDice-BA Argmax-prob
Closest papers by meaning
- On the Relaxation of Conditional Independence Assumption for Image Segmentation (NeurIPS 2026 · similarity 0.90)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.87)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.86)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.86)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.85)
- MaSS13K: A Matting-level Semantic Segmentation Benchmark (CVPR 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.