Addressing the Scale Shrinkage Problem in Learning-based Binocular Depth Estimation

Abstract

Binocular depth estimation is a fundamental problem in computer vision. Learning-based models have achieved significant performance improvements on public datasets in recent years. Our study finds that the performance of the current state-of-the-art deep learning-based models deteriorates significantly in distant areas. We point out that these deep learning-based models suffer from a scale shrinkage problem. Specifically, the predicted depth value ratio to the ground truth decreases as depth increases. Such a phenomenon is not conducive to the path planning and navigation of intelligent agents in outdoor scenes. We analyze the reasons for the scale shrinkage problem and give a simple and effective method. Our method employs a two-stage fine-tuning strategy and appropriately fuses the predictions of the two-stage models. The method does not reduce the prediction accuracy in close areas and significantly improves the accuracy of the models in distant areas. On the KITTI stereo 2015 dataset, our method can reduce the absolute relative difference by about 6% and the root-mean-square error (RMSE) by about 10%.