DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region

Jianping WuCVPR 2025 · DOI

Vanishing points (VPs)—the points where parallel lines in 3D space appear to converge in a 2D image—are fundamental to computer vision tasks like 3D reconstruction and autonomous driving. However, detecting them accurately and quickly has long been a tug-of-war between computational efficiency and mathematical complexity.

In the paper "DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region," Jianping Wu introduces a novel framework that breaks this deadlock. By mapping the infinite image plane onto a bounded disk and using a clever probabilistic partitioning scheme, DiskVPS achieves state-of-the-art accuracy at a blistering 158 frames per second.

The Challenge of Unbounded Space

The primary challenge in VP detection is the unboundedness of the image plane. Because parallel lines can converge anywhere—including at infinity—researchers must map the image plane into a bounded space to use practical techniques like the Hough Transform (HT).

Historically, there have been two main approaches:

  1. The Gaussian Sphere (GS): Maps the image to a unit sphere. While mathematically sound, it requires known camera calibration (focal length, optical center), making it useless for uncalibrated "in the wild" images.
  2. RANSAC-based methods: These work directly in the image plane but are often slow and prone to "false negatives" if the number of valid edge samples is small compared to the noise in a cluttered scene.

DiskVPS seeks a "third way": a method that is as robust as the Gaussian Sphere but requires zero calibration and runs fast enough for real-time applications.

Methodology

The core of DiskVPS is a three-step pipeline: Mapping, Partitioning, and Iterative Voting.

1. The Disk Mapping

The authors map the unbounded Cartesian coordinates $(u, v)$ of the image plane into a bounded polar coordinate system $(\psi, \theta)$. They use the transformation: $$\psi = \tan^{-1}(4r)$$ where $r$ is the distance from the image center. This elegantly squashes the entire infinite plane into a disk with a radius of $\pi/2$. This transformation ensures that even parallel lines converging at infinity are represented within the bounded region.

Mapping of parallel lines from Cartesian space (a) to the bounded disk region (b).

2. Equal-Probability Partitioning

A standard grid in a disk would create cells of wildly different "importance." Instead, the authors derived a statistical model to determine the probability of a random edge's extension intersecting any given point in the disk. They then partition the disk into concentric rings and fan-shaped cells such that each cell has an approximately equal probability of receiving a vote. This prevents discretization errors from biasing the detector toward certain regions of the image.

3. Individual Edge Voting

Unlike many HT methods that require pairs of edges to define a point (an $O(N^2)$ operation), DiskVPS allows individual edges to vote for all cells their extension line passes through. This reduces the complexity significantly. To find multiple VPs (like the three orthogonal VPs in a "Manhattan World"), the algorithm uses an iterative process: find the highest peak in the vote count, identify its inlier edges, remove their votes, and repeat.

Edge detection in an indoor scene (a) and the corresponding vote distribution in the Hough disk (b).

Key Results and Performance

DiskVPS sets a new bar for performance across major benchmarks like the York Urban Dataset (YUD), SVVP, and the synthetic SU3 dataset:

  • Speed: It achieves 158 FPS, making it roughly 4x faster than the previous fastest non-deep-learning SOTA (Wu et al., 39 FPS) and orders of magnitude faster than deep learning models like NeurVPS (0.5 FPS).
  • Accuracy: On the York Urban Dataset, DiskVPS achieved 74.5% accuracy at a 3° threshold, significantly outperforming the next best method (Tong et al. at 65.5%).
  • Robustness: The model showed very low sensitivity to changes in its internal parameters (the number of rings or cells), suggesting it is highly stable across different types of scenes.

Conclusion

DiskVPS achieves calibration-free accuracy comparable to or better than Gaussian Sphere methods without needing to know the camera's focal length. By using individual edges as voters and a pre-calculated probabilistic grid, the "heavy lifting" is done during the setup phase, allowing for real-time detection. While deep-learning-based models still hold a slight edge in pure angular accuracy on synthetic data, DiskVPS provides a highly efficient and versatile solution for real-world computer vision applications.

Summary updated

BibTeX
@inproceedings{wu2025diskvps,
  author = {Jianping Wu},
  title = {{DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02519},
}
Figure