Line Segment Detection: a Review of the 2022 State of the Art

We compare nine line segment detectors. The two more ancient ones are based on classical edge growing followed by a statistically founded validation. The next six are very recent and based on supervised deep learning. These six deep learning methods train and validate their neural network on two datasets (YorkUrban, Wireframe); most of them compared their results with the now classic LSD (Line Segment Detector) and EDlines, and get a better performance than them on these datasets. The ninth paper combines deep learning and classical edge growing to achieve a purely non-supervised method. The seven machine learning based detectors and EDlines are described here. LSD and EDlines are parameter-free, fixed to allow for one false alarm on average. Our experiments show that the six purely ML based line segment detectors show a significant variability to their end-parameters, leading to apparent missed or irrelevant detection. We also compared all nine detectors on two images: one clearly “in domain” for the Wireframe dataset, and the other one slightly out of domain. A quantitative comparison would be fallacious. Indeed, while differing in their search strategy, the statistical detectors share a very similar definition and decision threshold for line segments. The purely ML-based detectors have learned from human annotators that were directed at reconstructing architectures as wireframes. Hence, these algorithms aim at a different goal, the architectural interpretation of the scene. Yet, several of them have more complete goals than just line segment detection. Indeed, several of them also associate to each segment a descriptor, and aim at making the pair segment+descriptor fit for image matching. The readers are invited to judge by themselves about the advantages and drawbacks of all methods by submitting their own images to the online demos associated with the present paper.