en
0.25
0.5
0.75
1.25
1.5
1.75
2
A Naturalistic Open Source Movie for Optical Flow Evaluation
Published on Nov 12, 20124241 Views
Ground truth optical flow is difficult to measure in real scenes with natural motion. As a result, optical flow data sets are restricted in terms of size, complexity, and diversity, making optical flo
Related categories
Chapter list
A Naturalistic Open Source Movie for Optical Flow Evaluation00:00
Collaborators00:07
Advances driven by data (1)00:13
Advances driven by data (2)00:25
Middlebury Flow Dataset (2007) (1)00:29
Middlebury Flow Dataset (2007) (2)00:55
Error on Middlebury over time01:36
We need a challenging new dataset02:09
KITTI Vision Benchmark02:14
Introducing: MPI-Sintel03:18
Sintel: a Blender Open Movie03:52
Is synthetic data good enough?04:22
Idea: compare synthetic data to “lookalikes”04:33
Lookalikes04:39
Image statistics05:03
Image derivative log-histograms05:09
What about motion statistics?05:41
Flow statistics06:07
Speed histograms06:09
Realism story isn’t over06:42
CG data is not just “good enough”… … it has major advantages07:03
Render passes07:13
high flow gradient ^object boundaries07:40
Unmatched regions08:02
Results08:22
http://sintel.is.tue.mpg.de08:23
MDP-Flow2 estimated flow - MDP-Flow2 EPE09:06
Groundtruth - MDP-Flow2 EPE (1)09:18
Groundtruth - MDP-Flow2 EPE (2)09:26
Evaluation Take-aways09:34
Lessons learned10:07
Grand challenges for optical flow10:26
MJB and GBS were supported in part by NSF CRCNS Grant IIS-0904630.11:02