Computer Vision News - August 2017

4 / 56

TransFlow: Unsupervised Motion Flow…
Every month, Computer Vision News reviews a research from our field. This
month we have chosen to review two papers. The first one is TransFlow:
Unsupervised Motion Flow by Joint Geometric and Pixel-level Estimation. We
are indebted to the authors (Stefano Alletto, Davide Abati, Simone Calderara,
Rita Cucchiara, for allowing us to use their images to illustrate this
review. Their work is .
Background, motivation and novelty:
In the last few years, there is a growing interest among computer vision and
machine learning researchers about autonomous and assisted driving. Optical
flow estimation is one of the most researched topics in this context, but remains
an open problem. The automotive context: large displacements, extreme
changes in lighting conditions and the automotive movement making objects’
unique motion patterns very difficult to disentangle -- make optical flow
particularly challenging. Moreover, solutions that use deep learning typically
require large annotated datasets, however pixel-level annotated datasets are
lacking in the automotive field.
TransFlow builds upon several previous deep learning methods: the Spatial
Transformer (ST), developed by the DeepMind team in 2015, based on
convolutional neural networks, this unit can be added to any network to
perform explicit spatial transformations of features. Spatial transformers
produce models that learn invariance to translation, scale, rotation and other
generic warping. FlowNet, developed by Fischer et al, is one of the first end-to-
end deep architectures for dense optical flow. Using a convolutional-
deconvolutional autoencoder FlowNet provide a solution to the problem posed
by the absence of large annotated datasets by synthesizing an image dataset
featuring random chairs flying over random landscapes. The information is first
spatially compressed in a convolutional encoder block and then refined and re-
expanded in a transposed-convolutional decoder block, in a mirror architecture.
The novelty and advantages of TransFlow, compared to existing studies
conducted so far, are: (a) Significantly better generalization capabilities
compared to supervised approaches; (b) a simple and fast solution; (c) an end-
to-end forward-only neural network; and (d) outperformed other recent
attempts at unsupervised optical flow estimation.
Method:
The input and output of the TransFlow method:
4 Research
Research
Panasonic Silicon Valley Laboratory sponsored this work in cooperation
with Unimore, the University of Modena and Reggio Emilia.
Imagelab Unimore is grateful to Panasonic for its generous sponsorship.