cvg/DeepLSD

What it solves

DeepLSD addresses the challenge of extracting high-precision line segments from "in-the-wild" images. It bridges the gap between deep learning methods, which are typically robust but less precise, and handcrafted detectors, which are accurate but often less robust to noise or complex scenes.

How it works

The system predicts line distance and angle fields from an image. By combining these deep image gradients with an optimization process, it can detect generic line segments and refine existing line segments from any other detector to improve their accuracy.

Who it’s for

This tool is designed for developers and researchers working on tasks that require high-precision geometric information from images, such as 3D reconstruction, visual localization, and homography estimation.

Highlights

• Hybrid Approach: Combines deep learning robustness with handcrafted detector accuracy.
• Line Refinement: Can optimize and refine line segments already detected by other tools.
• Pre-trained Models: Includes models trained on Wireframe (for indoor scenes) and MegaDepth (for generic outdoor scenes).
• Evaluation Suite: Provides built-in tools for evaluating low-level line detection and vanishing point estimation across multiple datasets.