Recurrence-based Vanishing Point Detection

TL;DR

This paper addresses vanishing point detection in images lacking explicit straight lines by introducing Recurrence-based Vanishing Point Detection (R-VPD), an unsupervised method that fuses implicit lines from recurring patterns with explicit lines. The approach combines URP-based correspondences, hierarchical clustering, forward feature selection with geometric consistency scores, and an implicit-line fitting stage with a weighted RANSAC that culminates in a VP via eigen-decomposition. Two RP-based VP datasets are introduced: RPVP-Synthetic (3,200 images with ground-truth VP and camera parameters) and RPVP-Real (1,400 real-world images with human-annotated VP), enabling rigorous benchmarking against classical LSD/J-Linkage and deep-learning methods (NeurVPS, GPVPD). Results show R-VPD outperforms all baselines on the synthetic RPVP dataset and rivals state-of-the-art methods on real-world RP images, while maintaining competitive performance on broader TMM17 data; the work also discusses speed and memory trade-offs. The proposed unsupervised RP-driven framework broadens VP detection applicability, particularly in scenes with few explicit lines, with potential impact on structure-from-motion, camera calibration, and scene understanding in diverse environments.

Abstract

Classical approaches to Vanishing Point Detection (VPD) rely solely on the presence of explicit straight lines in images, while recent supervised deep learning approaches need labeled datasets for training. We propose an alternative unsupervised approach: Recurrence-based Vanishing Point Detection (R-VPD) that uses implicit lines discovered from recurring correspondences in addition to explicit lines. Furthermore, we contribute two Recurring-Pattern-for-Vanishing-Point (RPVP) datasets: 1) a Synthetic Image dataset with 3,200 ground truth vanishing points and camera parameters, and 2) a Real-World Image dataset with 1,400 human annotated vanishing points. We compare our method with two classical methods and two state-of-the-art deep learning-based VPD methods. We demonstrate that our unsupervised approach outperforms all the methods on the synthetic images dataset, outperforms the classical methods, and is on par with the supervised learning approaches on real-world images.

Figures (19)

• Figure 1: Two images containing examples of " things that recur" (Recurring Patterns). Top row: images with ground truth VP ($\times$) indicated. Bottom row: VP prediction results from four VPD methods R-VPD (our method), NeurVPS zhou2019neurvps, GPVPD lin2022deep and J-Linkage Toldo2008RobustMS respectively, where VP predictions that are too far from the ground truth are not shown in the images ((a) 2 misses, (b) 3 misses). Due to a lack of explicit straight lines, automatic vanishing point detection on these images poses challenges to explicit line-based methods.
• Figure 2: Recurrence-based Vanishing Point Detection: An overview of R-VPD is presented in the above image. SIFT features are extracted from the input image and are clustered hierarchically. Forward selection of the features is performed using geometric constraints based on linearity, angle, and scale. Implicit lines are fitted to the selected feature clusters, and a weighted RANSAC algorithm is employed to find the intersection of both implicit and explicit lines to locate the vanishing point. For this image, there are no explicit long lines.
• Figure 3: A Unit Recurring Pattern (URP) is an RP with at least two RP instances (RPIs) and each RPI has at least two distinct visual words (e.g. clustered SIFT descriptors) that correspond across the RPIs (this figure is adapted from zhang2022novel ). Here we add the two blue dotted lines passing through pairs of corresponding features, $(f_{11},f_{12})$ and $(f_{21},f_{22})$ , whose intersections identify the VP.
• Figure 4: Sample images from RPVP-Synthetic image dataset showing scenes in which RPs are oriented towards the VP. A relatively lesser number of explicit lines makes this dataset challenging for VPD. The "$\times$" represents the ground truth VP.
• Figure 5: The relationship between RPVP-Real, TMM17, RPVP-Real Exclusive and TMM17-Test. TMM17/TMM17-Test contains natural images with/without RPs while RPVP-Real/RPVP-Real Exclusive contains images with RPs only.