Authors:
Taha Alhersh
and
Heiner Stuckenschmidt
Affiliation:
Data and Web Science Group, University of Mannheim, 68131 Mannheim and Germany
Keyword(s):
Optical Flow, Unsupervised Learning, Fine-tuning, Motion Boundary, Deep Learning.
Related
Ontology
Subjects/Areas/Topics:
Computer Vision, Visualization and Computer Graphics
;
Motion, Tracking and Stereo Vision
;
Optical Flow and Motion Analyses
Abstract:
Recently, convolutional neural network (CNN) based approaches have proven to be successful in optical flow estimation in the supervised as well as in the unsupervised training paradigms. Supervised training requires large amounts of training data with task specific motion statistics. Usually, synthetic datasets are used for this purpose. Fully unsupervised approaches are usually harder to train and show weaker performance, although they have access to the true data statistics during training. In this paper we exploit a well-performing pre-trained model and fine-tune it in an unsupervised way using classical optical flow training objectives to learn the dataset specific statistics. Thus, per dataset training time can be reduced from days to less than 1 minute. Specifically, motion boundaries estimated by gradients in the optical flow field can be greatly improved using the proposed unsupervised fine-tuning.