Research map / Detection and segmentation / CNN features
DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region
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 DiskVPS, a non-learning vanishing point detector that maps the image plane into a disk, partitions it into equal-probability Hough cells, and votes with individual edges, outperforming SOTA on real-world YUD and SVVP benchmarks at high speed.
Problem
Vanishing point detection in uncalibrated images requires mapping the unbounded image plane into a bounded space for Hough transform; existing Gaussian-sphere methods need calibration and suffer from spurious detections, while deep learning methods need costly training data.
Approach
Map the image plane into a disk via tan^-1(4r) radial mapping, derive the probability distribution of edge-line intersections analytically, and partition the disk into concentric rings and fan-shaped cells of approximately equal probability so that individual edges (not edge pairs) can vote, with iterative inlier removal for VP extraction.
Key results
- On YUD, DiskVPS reaches 74.5% AA@3°, about 9 points above the second-best 65.5% (Tong et al. [27]), and 99.0% AA@10°.
- On SVVP, DiskVPS reaches 98.4% AA@10° and 81.6% AA@5°, while on synthetic SU3 it is competitive (84.3% AA@3°) but behind deep-learning methods.
- DiskVPS runs at 158 fps (6.3 ms per frame) on an Intel i7 3.5 GHz CPU, around 4x faster than the second-best at 39 fps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiskVPS instead of Tong et al. | VP detection · YUD | AA@3°: 65.5 -> 74.5 | Helped |
| use DiskVPS instead of Tong et al. | VP detection · YUD | AA@5°: 77.1 -> 90.2 | Helped |
| use DiskVPS instead of Tong et al. | VP detection · YUD | AA@10°: 87.4 -> 99.0 | Helped |
| use DiskVPS instead of Lin et al. | VP detection · SVVP | AA@3°: 51.8 -> 58.3 | Helped |
| use DiskVPS instead of Lin et al. | VP detection · SVVP | AA@5°: 68.7 -> 81.6 | Helped |
| use DiskVPS instead of Tong et al. | VP detection · SVVP | AA@10°: 82.6 -> 98.4 | Helped |
| use DiskVPS instead of NeurVPS | VP detection · SU3 | AA@3°: 93.9 -> 84.3 | Hurt |
| use DiskVPS instead of NeurVPS | VP detection · SU3 | AA@5°: 96.3 -> 91.7 | Hurt |
| use DiskVPS instead of Wu et al. | VP detection · YUD, SVVP, SU3 | FPS: 39 -> 158 | Helped |
| vary M from 200 to 300 and N from 100 to 180 (ablation) | DiskVPS · YUD | AA@3°: 70.6 to 78.4; mean 75.3; STDEV 2.2 | No clear effect |
| vary M from 200 to 300 and N from 100 to 180 (ablation) | DiskVPS · YUD | AA@5°: 89.2 to 94.1; mean 91.5; STDEV 1.3 | No clear effect |
| vary M from 200 to 300 and N from 100 to 180 (ablation) | DiskVPS · YUD | AA@10°: 99.0 to 99.0; mean 99.0; STDEV 0 | No clear effect |
Surprising findings
- On the synthetic SU3 dataset, the non-learning DiskVPS is outperformed by two deep-learning methods (NeurVPS and Tong et al.), which the authors attribute to SU3's large training size.
Limitations
- Relies on approximations of the PDF (e.g. grad(r) approximating frad, and 1/(pi*(4r)^2+1) for r>=1.5), so partitioning is only approximately equal-probability.
- Accuracy lags deep-learning methods on the large synthetic SU3 dataset.
- Evaluated only on Manhattan-world datasets with a fixed parameter setting; one tested zenith VP was placed at the opposite pole.
Experimental setup
Datasets: York Urban Dataset (YUD) SVVP SU3
Benchmarks: York Urban Dataset (YUD) SVVP SU3
Metrics: angular accuracy AA@3° AA@5° AA@10° frames per second
Baselines: J-Linkage Wu et al. [34] CONSAC NeurVPS Lin et al. [16] Tong et al. [27]
Builds on: Hough Transform Gaussian sphere Wu et al. probabilistic line intersection model [34]
Closest papers by meaning
- End-to-End Real-Time Vanishing Point Detection with Transformer (AAAI 2024 · similarity 0.85)
- VPDETR: End-to-End Vanishing Point DEtection TRansformers (AAAI 2024 · similarity 0.84)
- Convex Relaxation for Robust Vanishing Point Estimation in Manhattan World (CVPR 2025 · similarity 0.83)
- Selective Disk Bispectrum: A Complete and Rotation Invariant Image Descriptor (NeurIPS 2026 · similarity 0.82)
- Spatial Transform Decoupling for Oriented Object Detection (AAAI 2024 · similarity 0.80)
- V-DETR: DETR with Vertex Relative Position Encoding for 3D Object Detection (ICLR 2024 · similarity 0.79)
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.