Motion Models that Only Work Sometimes


It is too often that tracking algorithms lose track of interest points in image sequences. This persistent problem is difficult because the pixels around an interest point change in appearance or move in unpredictable ways. In this project we explore how classifying videos into categories of camera motion improves the tracking of interest points, by selecting the right specialist motion model for each video. As a proof of concept, we enumerate a small set of simple categories of camera motion and implement their corresponding specialized motion models. We evaluate the strategy of predicting the most appropriate motion model for each test sequence. Within the framework of a standard Bayesian tracking formulation, we compare this strategy to two standard motion models. Our tests on challenging real-world sequences show a significant improvement in tracking robustness, achieved with different kinds of supervision at training time.


BMVC 2012
Motion Models that Only Work Sometimes
Cristina García Cifuentes, Marc Sturzel, Frédéric Jurie and Gabriel J. Brostow.
BMVC 2012
[poster] [paper] [suppl. material] [video] [BibTeX]

Data and Code

If you find the following useful, please cite our BMVC paper (a journal version is in submission).

Motion Model Selection Dataset

The videos are in avi format, compressed with HuffYUV. They are separated into folders accoding 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].

Matlab code for recreating the experiments

Contact: cristina.garcia_cifuentes.09 (at)

Last updated: June 2013