Computer Vision News - August 2017

8 / 56

The authors and ourselves view the following as the most notable results of
TransFlow: the 3x performance improvement on unseen data, which
demonstrates that the method learns important transferable representations to
compute the flow. It proves its great generalizes capabilities - where hand tuned
methods had over fit to the dataset.
In the table above, Accuracy@5 is the ratio of motion vectors with end point
error lower than 5 pixels; APE is the average point error of all motion vectors.
8 Research
Research
“The most notable results of TransFlow are the 3x
performance improvement on unseen data, which
demonstrates that the method learns important
transferable representations to compute the flow”
“Lack of generality mainly due to
the synthetic nature of the rendered scenes”