SMV: Sequences with Motion Vectors
Sequences with Motion Vectors (SMV)
Overview
Motion compensation is an essential method in all modern video compression solutions, including hybrid video coding and neural-network-based video compression. As the true motion of objects in a given video is usually unknown, state-of-the-art encoders perform complex search methods to determine motion vectors suitable for efficient compression.
This dataset provides a set of computer-rendered videos with different motion characteristics. Next to the videos renderes by Blender, images are provided which represent the 2D-projected motion of objects on a per-pixel basis. This motion information can be used to speed-up or to enhance motion estimation.
Database
The database provides eleven scenes at CIF and at HD resolution with 15 up to 250 frames.
 arrowscif.zip |
 arrowshd.zip |
 hexacif.zip |
 hexahd.zip |
 hexastand.zip |
 hexacyc.zip |
 countrysoldier.zip |
 carchecker.zip |
 carshow.zip |
 pilotjumpcif.zip |
 pilotjumphd.zip |
Usage
Each sequence contains the following information:
Raw sequence in YUV420 format,
Motion and depth information for each frame in TIFF-format (with respect to the previous frame),
A cfg-file containing sequence information in the HM-configuration file format.
In the TIFF-files the information is stored in the following format:
R-channel: x-motion
G-channel: y-mtion
B-channel: z-motion (in depth domain, normalized to the range [0, 1])
Alpha-channel: depth (in the same domain as the z-motion)
Each channel is saved in 16 bits bit depth. If r and g are the color values of one image pixel (in the range [0, 2^16]), the corresponding motion in pixel domain (x,y) is calculated by
and
where w and h are the frame width and height, respectively. The normalization to the frame dimensions ensures that the maximum MV length can always be obtained. For all sequences, the motion information keeps quarter-pel accuracy, which is the maximum accuracy available in HEVC.
In C++, the tiff-files can be read using the freely available libtiff library.
Publication
If you use the dataset for your research, you should cite the follwing paper:
Christian Herglotz, David Müller, Andreas Weinlich, Frank Bauer, Michael Ortner, Marc Stamminger, André Kaup
Improving HEVC Encoding of Rendered Video Data Using True Motion Information
Proc. IEEE International Symposium on Multimedia (ISM), Taichung, Taiwan, Dec. 2018, pp. 287-290
DOI: 10.1109/ISM.2018.00063
arxiv: https://arxiv.org/abs/2309.06945
License
The database is licensed using CC-BY-SA.
Contact
lms-datasets@fau.de