A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation

TLDR

Optical flow estimation has been successfully framed as a supervised learning problem solved by convolutional networks, thanks to large synthetic datasets such as FlowNet. The study aims to extend convolutional network approaches from optical flow to disparity and scene flow estimation, introducing new datasets and a real‑time disparity network. The authors create three large synthetic stereo video datasets and train convolutional networks for disparity and joint flow‑disparity (scene flow) estimation, achieving real‑time performance. The datasets enable the first large‑scale training and evaluation of scene flow methods, and the joint network achieves the first scene flow estimation with a convolutional network.

Abstract

Recent work has shown that optical flow estimation can be formulated as a supervised learning task and can be successfully solved with convolutional networks. Training of the so-called FlowNet was enabled by a large synthetically generated dataset. The present paper extends the concept of optical flow estimation via convolutional networks to disparity and scene flow estimation. To this end, we propose three synthetic stereo video datasets with sufficient realism, variation, and size to successfully train large networks. Our datasets are the first large-scale datasets to enable training and evaluating scene flow methods. Besides the datasets, we present a convolutional network for real-time disparity estimation that provides state-of-the-art results. By combining a flow and disparity estimation network and training it jointly, we demonstrate the first scene flow estimation with a convolutional network.