A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation
Authors:
Nikolaus Mayer,
Eddy Ilg,
Philip Häusser,
Philipp Fischer,
Daniel Cremers,
Alexey Dosovitskiy,
Thomas Brox
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 pro…
▽ More
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.
△ Less
Submitted 7 December, 2015;
originally announced December 2015.
