The data is split into train and test subdirectories, each with a set of .txt files, one .txt file for each video clip (about 90% of the clips are in train, and the remaining 10% in test). The format of each .txt file is as follows:

  <Video URL>
  <frame1>
  <frame2>
  <...>
where each frame line has the following 19 columns:
     1. timestamp (int: microseconds since start of video)

   2-6. camera intrinsics (float: focal_length_x, focal_length_y,
                                  principal_point_x, principal_point_y)

  7-19. camera pose (floats forming 3x4 matrix in row-major order)
The camera intrinsics can be organized into a 3x3 matrix K and the camera pose parameters into a 3x4 matrix P = [ R | t ], such that the matrix KP maps a (homogeneous) 3D point p in a world coordinate frame to a (homogeneous) 2D point in the image.

The camera intrinsics are expressed in resolution-independent normalized image coordinates, where the top left corner of the image is (0,0), and the bottom right corner of the image is (1,1). This allows for the intrinsic parameters to be applied to frames at whatever resolution they are represented on disk (or resized to prior to training), by scaling them according to the image size in pixels. For an image of resolution width x height pixels, the intrinsics matrix at the actual scale of the image is


K =  
width · focal_length_x	0	width · principal_point_x
0	height · focal_length_y	height · principal_point_y
0	0	1
