PLB-Net: A Multi-Scale Attention-Based Network for Binocular Depth Estimation with Partial Labels

Vision-based depth estimation is increasingly recognized as a vital technique for applications in autonomous driving and robotics due to its cost-effectiveness compared to sensor-based methods. Among the various approaches within this domain, binocular depth estimation stands out due to its ability to mimic human stereoscopic vision by deriving depth from the disparity between images captured by two aligned cameras. Despite its advantages, this technique requires extensive labeling of data, which can be both time-consuming and costly. The urgency to address this challenge is evident, as existing methods largely neglect the practical constraints associated with partial label availability. The oversight in current research has motivated our development of PLB-Net, a novel convolutional neural network specifically designed to perform depth estimation using only partially labeled datasets. By integrating attention mechanisms, multi-scale feature learning, and adaptive feature extraction, PLB-Net significantly reduces the dependency on fully labeled data while maintaining high performance. Extensive experiments demonstrate that PLB-Net effectively addresses these challenges, confirming its robustness and utility in practical scenarios. This innovation makes PLB-Net especially suited for real-world applications where acquiring extensive ground truth labels is impractical, marking it as a pivotal development in the evolution of depth estimation technologies.