STTR-3D: Stereo Transformer 3D Network for Video-Based Disparity Change Estimation

In the field of computer vision and stereo depth estimation, there has been little research in obtaining high-accuracy disparity change maps from two-dimensional images. This map offers information that fills the gap between optical flow and depth which is desirable for numerous academic research problems and industrial applications, such as navigation systems, driving assistance, and autonomous systems. We introduce STTR3D, a 3D extension of the STereo TRansformer (STTR) which leverages transformers and an attention mechanism to handle stereo depth estimation. We further make use of the Scene Flow Flying-Things-3D dataset which openly includes data for disparity change and apply 1) refinements through the use of MLP over relative position encoding and 2) regression head with an entropy-regularized optimal transport to obtain a disparity change map. This model consistently demonstrates superior performance for depth estimation tasks as compared to the original model. Compared to the existing supervised learning methods for estimating stereo depth, our technique simultaneously handles disparity estimation and the disparity change problem with an end-to-end network, also establishing that the addition of our transformer yields improved performance that achieves high precision for both issues.