Disp1DNet: Rapid 1D Stereo Network for Efficient Disparity Map Estimation

One of the significant challenges in the field of computer vision is the calculation of scene depth and the distance of objects from images obtained by cameras. In computer vision, image depth is derived using various methods, one of which is stereo vision. In this approach, the disparity of objects in images seen by two eyes serves as the basis for calculating distance and scene depth. An important aspect of this discussion is the ability to process in real-time and beyond real-time since subsequent stages require considerable time for post-processing and decision-making. Additionally, in various applications, such as mobile phones, size, weight, power consumption, and hardware costs become crucial factors. Thus, having a model that not only offers good performance but can also be implemented on low-power hardware becomes particularly important. There are various methods for stereo vision calculations, generally categorized into traditional methods and neural network-based methods. Among these two categories, neural network-based models provide higher accuracy results compared to traditional methods, which has motivated researchers in recent years to present diverse models capable of real-time execution. After examining the foundations of the stereo vision domain and reviewing existing methods, we present a creative comprehensive model called Disp1DNet in this paper. Unlike existing models that process entire images using two-dimensional and three-dimensional convolutional layers, this model utilizes one-dimensional convolution and processes images row by row. This innovation allows us to achieve a processing time of 1.6 ms with a 22% error rate on the Synthetic Stereo Image Data for Algorithm Evaluation dataset, which relates to driving, approximately 20 times less than the minimum time required for real-time processing.