Abstract
We propose an end-to-end method for simultaneously detecting local junctions and global wireframe in man-made environment. Our pipeline consists of an anchor-free junction detection module, a distance map learning module, and a line segment proposing and verification module. A set of line segments are proposed from the predicted junctions with guidance of the learned distance map, and further verified by the proposal verification module. Experimental results show that our method outperforms the previous state-of-the-art wireframe parser by a descent margin. In terms of line segments detection, our method shows competitive performance on standard benchmarks. The proposed networks are end-to-end trainable and efficient.
Introduction
Inferring the 3D structure of a scene from a single 2D image is a fundamental challenge in computer vision. In man-made environments, such as indoor offices or urban streets, traditional local features often struggle due to textureless surfaces or repetitive patterns. Wireframe parsing addresses this by extracting a structural "skeleton" composed of salient junctions and the lines connecting them.
Previous approaches faced significant hurdles, including rigid, anchor-based junction detection that failed in dense areas and a lack of robust verification for proposed line segments, often resulting in "ghost" connections. This paper introduces an end-to-end framework that overcomes these limitations using a novel distance map learning module.
Proposed Method
The proposed pipeline is a two-stage, end-to-end trainable network:
1. Feature Extraction and Distance Map Learning
The network utilizes a "Stacked Hourglass" architecture to generate three primary per-pixel maps: a Junction Heatmap, a Junction Branch Map, and a Distance Map. The Distance Map is the core innovation; each pixel value represents its distance to the nearest line segment, providing a continuous, global field that guides the parsing process.
2. Line Proposing and Verification
Instead of relying solely on predicted angles, the system employs a "Least Distance" algorithm. It uses the learned distance map to evaluate potential line segments by checking the "summed distance" along the path. Candidate lines are then passed through a modified RoI-Align module and a classification network for final verification.
Key Results
The model was evaluated on the Wireframe Dataset and the York Urban Dataset.
- Accuracy: The method outperformed previous state-of-the-art (SOTA) models on the Wireframe dataset by a notable margin in precision and recall.
- Efficiency: Achieving approximately 6.7 FPS on an NVIDIA Titan X, the system is over three times faster than the previous leading method (2.0 FPS).
- Qualitative Quality: The resulting wireframes are visually cleaner, handling cluttered environments with fewer false connections.
Conclusion
This work demonstrates that incorporating global geometric cues via distance maps significantly enhances wireframe parsing. Future work may involve utilizing large-scale synthetic data to learn even more complex geometric relationships and further optimizing the architecture for real-time robotics applications.
Summary updated