Abstract
Dynamic textures (DT) are typically 3D videos of physical processes showing statistical regularity but have indeterminate spatial and temporal extent. Existing DT recognition methods usually neglect the global spatio-temporal relationships of DT which reflect the statistical regularities. In this paper, a spatio-temporal texture convolutional neural network (SMTCNN) is proposed for global semantic DT representation. Specifically, SMTCNN describes DT features by learning DT’ temporal motion as well as the sources of the motions and the scenarios where the motion is happening, and accordingly, a motion net and a source net are formulated. In particular, a novel module consisting of expansion and concatenation implementations on deep features is presented, with an arbitrary 2D backbone as input, followed by a new 1D CNN including 4 convolutional, 2 pooling and 2 fully-connected layers to represent the 2D tensors in space–time, by transforming DT descriptors from discrete “words” to global “textures”. A number of comparative experiments on three DT dataset - UCLA, DynTex and DynTex++ are conducted to demonstrate our approach.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have