Enhancing stereo vision estimation with custom Berhu-Gradient loss

Abstract

Stereo vision and disparity estimation play a critical role in allowing machines to perceive depth, with its applications spanning from various real-time systems such as autonomous driving, robotics to medical imaging. The practicality of these algorithms in real-time systems are hence tied closely to its efficiency and accuracy. As such, exploration have been done to employ deep learning for stereo vision, aimed at improving the performance in terms of accuracy and efficiency. Deep learning models generally outperform traditional stereo vision techniques in handling challenging regions like occlusions and texture-less regions. Despite these advancements, optimising disparity estimation models for higher accuracy and efficiency is still an ongoing challenge. Thus, we would like to explore a novel loss function and evaluate its effects on the accuracy of disparity estimation. This research aims to analyse the accuracy of the resulting model from our novel loss function. To achieve this, we train a baseline model with Huber loss function and compare it against our model which is trained on Berhu-Gradient loss function. The methodology involves modifying the loss function from a smooth L1 function (Huber Loss) to a hybrid between Berhu Loss and Gradient Loss. We will compare the results of the baseline model against our model and discuss the implications of the new loss function. The evaluation metric used includes end-point error (EPE) and disparity accuracy thresholds to assess improvements in depth estimation performance. To further improve on the accuracy, an ablation study is done to fine tune hyperparameters within the new hybrid function.