Abstract
Junctions play an important role in the characterization of local geometric structures in images, the detection of which is a longstanding and challenging task. Existing junction detectors usually focus on identifying the junction locations and the orientations of the junction branches while ignoring their scales; however, these scales also contain rich geometric information. This paper presents a novel approach to junction detection and characterization that exploits the locally anisotropic geometries of a junction and estimates the scales of these geometries using an a contrario model. The output junctions have anisotropic scales --- i.e., each branch of a junction is associated with an independent scale parameter --- and are thus termed anisotropic-scale junctions (ASJs). We then apply the newly detected ASJs for the matching of indoor images, in which there may be dramatic changes in viewpoint and the detected local visual features, e.g., key-points, are usually insufficiently distinctive. We propose to use the anisotropic geometries of our junctions to improve the matching precision for indoor images. Matching results obtained on sets of indoor images demonstrate that our approach achieves state-of-the-art performance in indoor image matching.
Introduction
Indoor image matching is a fundamental challenge in computer vision, particularly for tasks like 3D reconstruction and SLAM. Unlike outdoor scenes, indoor environments are often characterized by low-texture surfaces and repetitive geometric patterns. Standard feature descriptors like SIFT frequently fail in these settings because they rely on local photometric gradients that are often absent on blank walls or distorted by extreme viewpoint changes.
The core issue with existing junction detectors is their "isotropic" nature. They typically analyze a fixed circular region around a point, ignoring the varying lengths of the lines (branches) that form the junction. This paper proposes Anisotropic-Scale Junctions (ASJs), which treat junctions as structural anchors where each branch has its own independent scale, capturing the "geometric backbone" of a scene.
Methods
The proposed framework detects and matches ASJs through several key innovations:
- Anisotropic Modeling: Instead of a single radius for all branches, the ASJ model assigns a specific scale ($r$) and orientation ($\theta$) to each branch. This captures the inherent asymmetry of real-world architecture, such as where a short window frame meets a long ceiling line.
- A-Contrario Scale Estimation: To robustly determine branch lengths, the authors employ an a-contrario mathematical framework. By calculating the Number of False Alarms (NFA), the system identifies "meaningful" branches that are statistically unlikely to occur by chance in random noise, allowing for automatic thresholding without manual tuning.
- Differential Junction-ness: A differential model evaluates how "junction-like" a region remains as scale increases, optimizing the endpoint detection for long architectural lines common in indoor environments.
- Geometric Matching: An ASJ with at least two branches provides three distinct coordinates (the center and two endpoints). This configuration allows for the direct estimation of a local homography (an affine transform), which is significantly more effective at handling perspective shifts than single-point matching.
Results and Key Insights
The ASJ method was evaluated on over 100 indoor image pairs, including scenes with severe viewpoint changes and minimal texture.
The quantitative results demonstrate a clear advantage over traditional methods:
- Matching Accuracy: ASJs achieved 85.17% accuracy, significantly higher than SIFT (62.69%) and Affine-SIFT (82.85%).
- Structural Integrity: Unlike Affine-SIFT, which often clusters matches on small textured details, ASJ matches align with the primary geometric layout of the room.
- Line Segment Matching: ASJs also improved line matching accuracy to 71.55%, compared to 52.95% for the previous state-of-the-art (LJL).
Conclusion
By prioritizing geometry over photometry, the Anisotropic-Scale Junction approach provides a robust solution for indoor image correspondence. While the statistical estimation adds some computational complexity, the ability to derive precise local homographies from structural anchors represents a significant step forward for indoor scene analysis. Future work may involve integrating these junctions into global 3D reconstruction pipelines to further refine architectural modeling.
Summary updated