Research map / Detection and segmentation / CNN features
Hierarchical Histogram Threshold Segmentation - Auto-terminating High-detail Oversegmentation
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 HHTS, a top-down hierarchical superpixel method that recursively splits segments via multi-channel histogram thresholding and auto-terminates when color information is exhausted, reporting higher boundary recall on BSDS500 and NYUV2 and refining SAM masks.
Problem
Superpixel oversegmentation methods typically require predefined seed initialization and user-set termination criteria (superpixel count or iteration limit), which are hard to tune and can lose fine details, limiting boundary adherence.
Approach
Iteratively splits segments into foreground and background using local histogram thresholds chosen by Laplace-filter boundary detection with equal-partition weights, applied jointly across RGB, HSV and LAB channels, prioritizing segments by max per-channel standard deviation times size squared; it stops automatically when every segment has a single intensity per channel.
Key results
- On BSDS500 at 1000 superpixels, HHTS reaches BR 0.9626 and UE 0.0307 versus VSSS's BR 0.9188 and UE 0.0324.
- HHTS-preETPS (ETPS pre-segmentation refined by HHTS) is reported as 26.3% faster than pure HHTS on BSDS500 images.
- Average HHTS runtime is about 448.0 ms per BSDS500 image at 600 superpixels (single-threaded, CPU-only), yielding 172 levels per image on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HHTS instead of VSSS at 1000 superpixels | HHTS · BSDS500 | BR: 0.9188 -> 0.9626 | Helped |
| Replace the proposed histogram thresholding with OTSU thresholding (ablation) | HHTS · BSDS500 | BR | Hurt |
| Replace the proposed histogram thresholding with local histogram minima (ablation) | HHTS · BSDS500 | BR | Hurt |
| Disable equal partition weights (ablation) | HHTS · BSDS500 | BR | Hurt |
| Use RGB only instead of the full multi-channel approach (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Use HSV only instead of the full multi-channel approach (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Use LAB only instead of the full multi-channel approach (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Apply Gaussian blurring with a 3 × 3 kernel (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Use 16 histogram bins (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Use 64 histogram bins (ablation) | HHTS · BSDS500 | BR | No clear effect |
| Increase minimum detail size to 128 pixels (ablation) | HHTS · BSDS500 | BR | Helped |
| Reduce minimum detail size to 32 pixels (ablation) | HHTS · BSDS500 | BR | Mixed |
Surprising findings
- HHTS achieves lower compactness and boundary precision than most baselines, a trade-off attributed to irregular, elongated shapes.
Limitations
- Lower compactness and boundary precision due to irregular superpixel shapes.
- Early iterations can leak across ground-truth segments, increasing undersegmentation error.
- Runtime is not optimized; relies on local histogram and connected-component computation.
Experimental setup
Datasets: BSDS500 NYUV2
Benchmarks: BSDS500 NYUV2
Metrics: boundary recall undersegmentation error mean distance to edge achievable segmentation accuracy explained variation intra-cluster variation compactness boundary precision runtime
Baselines: ETPS SEEDS ERS CRS ERGC SLIC SH SCAC VSSS APENet LDFUNet CRTrees
Builds on: SLIC Otsu thresholding ETPS
Closest papers by meaning
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.84)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.84)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · similarity 0.84)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.84)
- BLO-SAM: Bi-level Optimization Based Finetuning of the Segment Anything Model for Overfitting-Preventing Semantic Segmentation (ICML 2024 · similarity 0.84)
- Segment Anything without Supervision (NeurIPS 2024 · similarity 0.84)
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.