|Description (include details on usage, files and paper references)||The UCL Motion Model Selection Dataset contains videos in avi format, compressed with HuffYUV. They are separated into folders according to manual inspection-based labels, as explained in the paper.
* Brownian videos [zip 498 MB]
* Constant Velocity videos [zip 372 MB]
* Traveling Right videos [zip 643 MB]
* Traveling Left videos [zip 752 MB]
* Forward videos [zip 3 GB]
* Backward videos [zip 692 MB]
Other related files:
* Dataset description [txt]
* Inspection labels [txt]
* Intro video [short: avi 232 MB] [full: mp4 280 MB]
We provide sparse ground-truth tracks for each clip [zip 1.1 MB]. They were obtained using our annotation tool [binaries and code]. Check its documentation for information about the output file format. This annotation tool is fed with FAST interest point detections at every frame, also provided [zip 68 MB].
Our dataset consists of over 100 challenging real-world videos from YouTube and Vimeo encompassing various scenarios, illuminations, and recording conditions. They last typically between 3 and 90 seconds.
To train a supervised classifier, we inspected the videos and judged whether a clip belonged predominantly to one of four rough camera motions (Traveling Right, Traveling Left, Forward or Backward), or might be better served by the Brownian or Constant Velocity models. For the TRight and TLeft categories, the paper also used the flipped video as part of the dataset. In total, there are 12 videos labeled as Br, 11 as CVel, 11 (+ 17 flips) as TRight, 17 (+ 11) as TLeft, 24 as Fwd, and 13 as Bwd.
In the BMVC paper, there is a 14th Bwd spurious video used by mistake. It has been removed from the dataset since.
All the videos were downloaded from the internet. Big Thank you! to the authors and uploaders.
Motion Models that Only Work Sometimes
Cristina García Cifuentes, Marc Sturzel, Frédéric Jurie and Gabriel J. Brostow.