MLSD Line Detection

Yes, MLSD stands for Mobile Line Segment Detection (M-LSD), developed by NAVER and LINE Vision. It is widely referenced as M-LSD in academic research and as the MLSD preprocessor in generative diffusion workflows.

MLSD is purpose-built for straight geometric geometry: interior room layouts, architectural building facades, street perspectives, furniture framing, and structural grids. It extracts crisp vector-like wireframes while stripping away surface textures, shadows, and color noise.

Avoid using MLSD for organic silhouettes, human figures, animals, curved sculptures, vegetation, and soft drapery. Because the detector exclusively tracks straight segments, organic forms will either be ignored or rendered as fragmented visual artifacts. For organic shapes, reach for edge or depth preprocessors instead.

Score Threshold sets the confidence cutoff between 0 and 1 (defaulting to 0.1) to eliminate noisy or faint line detections in busy textures. Distance Threshold spans from 0 to 20 (defaulting to 0.1) and controls line merging—lower values preserve fine parallel architectural details like window mullions, while higher values consolidate overlapping segments into simplified guides.

NAVER and LINE Vision introduced M-LSD at AAAI 2022. The model uses a streamlined MobileNetV2 backbone with roughly 25,000 parameters and a tri-points formulation to deliver real-time single-shot line detection on resource-constrained hardware.