
========================================================
Sequences for three-view perspective motion segmentation
========================================================

General information
-------------------

These sequences have been in [1] to test three-view perspective motion segmentation algorithms.
These sequence are a subset of the Hopkins 155 dataset [2]

File format
-----------

Each sequence is contained in a directory having the corresponding
name.  In each directory there are some files.
. description.txt: a brief description of what appears in the
  sequence.
. preview.jpg: the first frame of the sequences with the extracted
  feature points.
. *.avi and *_points.avi: the video sequence in the original form and
  with tracked correspondences superimposed.
. *_truth.mat: a Matlab 6 .MAT file containing the ground-truth.

The variables inside the ground-truth file are organized as follow:
. width and height: dimensions (in pixels) of all the frames in the
  video sequence.
. points: number of tracked points P.
. frames: number of frames F.
. y: a matrix 3xPxF containing the homogeneous coordinates of the P
  points in the F frames.
. x: a matrix 3xPxF derived from y by normalizing the first two
  components of each vector such that they belong to the interval
  [-1;1].
. K: the 3x3 normalization matrix used to pass from y to x
  (x=K^(-1)*x).
. s: a Px1 vector containing the ground-truth segmentation; for each
  point it gives the index of the corresponding motion group.

References
----------
[1] R. Vidal and R. Hartley.
"Three-View Multibody Structure from Motion"
IEEE Transactions on Pattern Analysis and Machine Intelligence, volume 30, number 2, pages 214 - 227, 2008. 

[2] http://www.vision.jhu.edu/data/hopkins155/
