LCLD: A lightweight vanishing point detector with contrast-learning-based intermediate supervision module

Vanishing point detection is crucial in 3D vision, enabling the extraction of 3D information from 2D images. However, many vanishing point detectors involve a trade-off between model complexity and detection accuracy. To address this problem, we propose a lightweight vanishing point detector with intermediate supervision and a classifier for channel aggregation (CIAP). The proposed approach has the following novelties. Firstly, the intermediate supervision module leverages contrast learning, which learns by bringing similar samples closer and pushing dissimilar ones apart, with an extremely positive and negative sample selection strategy. Secondly, the fully connected layers are replaced with purely convolutional layers that aggregate multi-channel information, reducing the model parameters from \(\varvec{22M}\) $\mathbf{22}\mathit{M}$ to \(\varvec{4.6M}\) $\mathbf{4.6}\mathit{M}$ without compromising accuracy. Extensive validation on synthetic and real-world datasets shows the strong performance of our approach, with a \(\varvec{5.4\%}\) $\mathbf{5.4}\mathbf{\%}$ improvement in angle accuracy \(\varvec{0.2^{\circ }}\) $\mathbf{0}\mathbf{.}{\mathbf{2}}^{\mathbf{\circ }}$ over the state-of-the-art method VaPiD on the synthetic dataset. The reduced parameter count supports energy-efficient systems, contributing to the development of sustainable and scalable 3D vision solutions.