Deep Learning for Optical Flow
Lead Research Organisation:
University of Southampton
Department Name: Electronics and Computer Science
Abstract
Optical flow is a classic problem of computer vision which has seen much and varied work
over the years. Recently, deep learning has revolutionised many image tasks that were
previously considered extremely difficult however it's application in video is less studied.
This work shall further explore the work done with the first deep models for optical flow
and propose 5 new models for it's estimation. It is found that 3D convolutions - over
space and time - perform better than naive methods for dealing with the temporal
aspect of the optical flow problem. It's further found that the usual starting point for
temporal problems - a recurrent network - performs much worse for this task and is
more computationally and memory intensive.
over the years. Recently, deep learning has revolutionised many image tasks that were
previously considered extremely difficult however it's application in video is less studied.
This work shall further explore the work done with the first deep models for optical flow
and propose 5 new models for it's estimation. It is found that 3D convolutions - over
space and time - perform better than naive methods for dealing with the temporal
aspect of the optical flow problem. It's further found that the usual starting point for
temporal problems - a recurrent network - performs much worse for this task and is
more computationally and memory intensive.
Organisations
People |
ORCID iD |
Adam Prugel-Bennett (Primary Supervisor) | |
Matthew Painter (Student) |
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509747/1 | 01/10/2016 | 30/09/2021 | |||
1952364 | Studentship | EP/N509747/1 | 01/10/2017 | 31/12/2020 | Matthew Painter |