Recurrence-Based Vanishing Point Detection

Skanda Bharadwaj, Robert T. Collins, Yanxi LiuWACV 2025

TL;DR

This paper introduces Recurrence-based Vanishing Point Detection (R-VPD), a novel unsupervised method that detects vanishing points (VPs) by leveraging "recurring patterns" (RPs) and their implicit line correspondences, rather than solely relying on explicit straight lines. This approach significantly improves VP detection in images lacking clear line segments, a common challenge for existing methods.

What Problem Does This Paper Address?

Detecting vanishing points (VPs) is crucial for many computer vision tasks like 3D reconstruction and camera calibration. Traditional methods often rely on identifying explicit straight line segments in an image and then clustering them to find VPs. However, these methods struggle in scenes where such explicit lines are scarce or absent, a common occurrence in real-world imagery. While recent deep learning approaches have shown promise, they typically require large, labeled datasets for training. This paper addresses the gap of robust, unsupervised VP detection in images that lack explicit straight lines, by proposing a method that utilizes implicit geometric cues derived from recurring patterns.

Methods

The core of R-VPD lies in its ability to discover and utilize "implicit lines" derived from recurring patterns (RPs). RPs are defined as objects or structures that frequently appear together in images, suggesting a shared geometric relationship. The method starts by extracting SIFT features and then uses a hierarchical clustering approach to group these features.

The key innovation is the discovery of "implicit lines" through feature correspondences within these recurring patterns. Instead of fitting lines to detected edges, R-VPD fits lines through pairs of corresponding features identified during the RP discovery process. These implicit lines, along with any explicit lines detected in the image, are then used in a weighted RANSAC algorithm to robustly estimate the vanishing point. The weighting in RANSAC is based on the angular relationships between lines, giving more importance to lines that are more likely to contribute to a consistent VP.

Datasets

The authors introduce two new benchmark datasets:

  • RPVP-Synthetic: A dataset of 3,200 synthetic images generated in Blender. It includes ground truth VP locations and camera parameters (intrinsic and extrinsic matrices). This dataset is designed to feature scenes with RPs but often lacks explicit straight lines.
  • RPVP-Real: A dataset of 1,400 real-world images with human-annotated vanishing points. These images also contain recurring patterns and may or may not have explicit straight lines.

They also use the TMM17-Test dataset, a standard benchmark for VP detection, for comparison.

Key Insights and Ideas

  • Implicit Lines from Recurring Patterns: The most significant insight is that recurring patterns, even without explicit straight lines, provide strong geometric cues. By finding correspondences within these patterns, implicit lines can be inferred, which are as valuable as explicit lines for VP detection.
  • Unsupervised Robustness: R-VPD's unsupervised nature makes it adaptable to various scenes without the need for extensive labeled data, a major advantage over supervised deep learning methods.
  • Complementary to Explicit Lines: The method effectively combines implicit lines with explicit lines, enhancing its robustness across a wider range of image conditions.
  • Challenging Existing Methods: The introduction of datasets with limited explicit lines highlights the limitations of traditional and even some state-of-the-art methods, demonstrating the need for approaches like R-VPD.

Key Results

  • Outperforms on Synthetic Data: R-VPD significantly outperforms classical and deep learning methods on the RPVP-Synthetic dataset, achieving superior accuracy metrics.
  • Competitive on Real-World Data: On the RPVP-Real dataset, R-VPD performs comparably to state-of-the-art deep learning methods, and significantly better than classical methods.
  • On Par with SOTA on Non-RP Data: Even on the TMM17-Test dataset (which doesn't exclusively feature RPs), R-VPD performs competitively, often second best, with statistically insignificant differences compared to the top performer.
  • Qualitative Success: The paper shows qualitative results where R-VPD successfully detects VPs in challenging images where other methods fail, particularly those lacking explicit straight lines.

Limitations and Future Work

  • Limitations on RP Images: The authors acknowledge that R-VPD has limitations on images with very strong perspective distortions and on images that do not contain recurring patterns.
  • Single Vanishing Point Focus: The current work focuses on scenes with a single vanishing point.
  • Future Directions: The authors aim to extend R-VPD to handle images with multiple vanishing points (e.g., Manhattan-world scenes) and to improve its robustness on general indoor and outdoor scenes, potentially by incorporating more sophisticated RP detection or by extending its applicability beyond images with single VPs.

Summary updated

BibTeX
@inproceedings{bharadwaj2025recurrence,
  author = {Skanda Bharadwaj and Robert T. Collins and Yanxi Liu},
  title = {{Recurrence-Based Vanishing Point Detection}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure