LSDNet: Trainable Modification of LSD Algorithm for Real-Time Line Segment Detection

Official implementation of "LSDNet: Trainable Modification of LSD Algorithm for Real-Time Line Segment Detection." by Teplyakov, Lev, Leonid Erlygin, and Evgeny Shvets. IEEE Access 10 (2022): 45256-45265.

links to the paper: IEEE Access, arXiv.org

Quantitative Results

Qualitative Results

Video demo

Images from Wireframe dataset of varying number of salient line segments

LSD detector

HAWP detector

M-LSD-tidy detector

LSDNet, the proposed detector

Setting up LSDNet Project

WARNING: all the scripts run in docker container without GPU support. We preferred plug-and-play demo with low FPS to top performance with hardcore NVIDIA-docker installation.

Installation

cd devops
sh build_image.sh

it takes about 10 min, it will download all the requirements and compile OpenCV the way we need it.

Launch Interactive Demo

The Interactive Demo allows you to play with LSDNet on single images.

To launch the demo server (based on Gradio) run

cd devops
sh launch_web_vis.sh

This commad will launch a web server and print its address. The interface should look like this.

Running on Your Images

To make predictions on your images, specify in devops/detect_lines.sh the following variables:

input_path - path to an images folder or a single image (images must be png or jpg)
out_path - path where predicted segments will be saved in .txt format (x1,y1, x2,y2,conf) - the coordinates of segments' endpoints and segment confidence

As a container mounts to the repo's root directory, paths above must be:

relative to lsd/detect_lines.py file
inside the repo's directory

Then make predictions on your images with

cd devops
sh detect_lines.sh