Self-supervised Lightweight Line Segment Detector and Descriptor

Image line segment features can be used as a complement to points for geometric computer vision tasks. Previous deep learning-based line segment feature extraction is difficult to perform real-time inference in environments with limited computational resources due to the huge model size and computational cost problems. In this paper, we introduce the joint detection and description of line segments using lightweight networks in resource-constrained environments. We use a lightweight backbone as an encoder and design a top-down architecture to fuse multi-layer features, which is split into three branches for different tasks. Moreover, thanks to the self-supervised training scheme, our model can be trained on dataset without ground-truth line segments. In the experiments, our approach achieves competitive performance with the fastest model inference speed.