The Impact of Semi-Supervised Learning on Line Segment Detection

TL;DR

This paper presents a method, to their knowledge the first to target line detection using modern state-of-the-art methodologies for semi-supervised learning, based on a semi-supervised framework and shows comparable results to fully supervised methods.

Abstract

In this paper we present a method for line segment detection in images, based on a semi-supervised framework. Leveraging the use of a consistency loss based on differently augmented and perturbed unlabeled images with a small amount of labeled data, we show comparable results to fully supervised methods. This opens up application scenarios where annotation is difficult or expensive, and for domain specific adaptation of models. We are specifically interested in real-time and online applications, and investigate small and efficient learning backbones. Our method is to our knowledge the first to target line detection using modern state-of-the-art methodologies for semi-supervised learning. We test the method on both standard benchmarks and domain specific scenarios for forestry applications, showing the tractability of the proposed method.

Figures (6)

• Figure 1: From left to right: Ground truth detection, supervised model on 1/16 of data, supervised model on all data, proposed semi-supervised approach with 1/16 labeled data. The semi-supervised method improves significantly on the supervised method.
• Figure 2: Many trained models fail to generalize to new image domains. Top row shows the output from left to right: DeepLSD pautrat2023deeplsd, LETR xu2021line, M-LSD mlsd, trained on the Wireframe dataset. Bottom row shows the output from these models on an out of domain image.
• Figure 3: System overview of our proposed method. Here $\operatorname{F }$ denotes the shared model, $x$ is labeled data, $x^w$ weakly perturbed unlabeled data, and $x^{s_{1,2}}$ strongly perturbed data. The loss function for the labeled data, $\mathcal{L}_{labeled}$ , is described in section \ref{['sec:labeled']}, and is directly built on the loss from mlsd. For the unlabeled stream, $\operatorname{H}$ denotes the cross-entropy loss described in section \ref{['sec:unlabeled']} and is calculated between the weakly perturbed sample and one of the strongly perturbed samples.
• Figure 4: Comparison of the output results on one image from the Finnwoodlands test dataset. From left to right, the fully supervised model, our proposed method trained with $1/2$ the dataset labeled, and the ground truth. Both models are trained on the Finnwoodlands train dataset.
• Figure 5: Comparison of the output results on one image from the Wireframe test dataset. From left to right, the supervised model trained with $1/16$ the dataset labeled, our proposed method trained with $1/16$ the dataset labeled and the remaining as unlabeled, and the ground truth.