Each of the following datasets contain a series of RGB and depth images of a planar scene. In each dataset, the camera performs a distinct movement.

The images were captured using a Kinect camera and have a resolution of 1280x960 (the depth images are upscaled). They were generated using the ROS-based rgbd_dataset_gen toolkit.

The data sets are 'downwards compatible' to the image datasets created by Mikolajczyk et. al.

David Gossow, David Weikersdorfer, Michael Beetz: Distinctive Texture Features from Perspective-Invariant Keypoints, International Conference on Pattern Recognition (ICPR), 2012

You can download datasets referenced in the publication in one single file: daft_datasets.tar.gz (207 MB)

You can get the source code of DAFT on github: https://github.com/dgossow/daft.

You can find a further list of related publications on the site of the original evaluation framework: http://www.robots.ox.ac.uk/~vgg/research/affine/.

Each directory contains the following files:

• depth*.pgm: 16-Bit depth images in mm
• img*.pgm: 8-Bit RGB images
• mask.pgm: Indicates region of interest in first image
• H1to*p: 3x3 Homography from image 1 to n
• K: 3x3 intrinsic camera matrix

In addition, it contains one of the following parameter files:

• cam_rotation: rotation around camera axis (degrees)
• vp_rotation: rotation around plane normal (degrees)
• vp_angle: angle between camera axis and plane normal, relative to img1 (degrees)
• scaling: 2d image scaling relative to first image (0..1)

The parameter files contain the numerical values for all images except the first.

Download the full dataset: poster.tar.gz (623 MB)

This scene shows several posters fixed to the ground or wall.

It contains the following transformations:

• camrotate0,camrotate45,camrotate60: The camera rotates up to 90° around its view axis, while looking at the posters in a fronto-parallel view or 45°/60° tilted to the side.

• scale: The camera moves away from the scene, resulting in a scaling effect down to 24%.

• vpangle0, vpangle22, vpangle45: The camera starts at a fronto-parallel view, tilting to the side relative to the plane. The optical tilt axis is at 0, 22 and 45 degrees.

• vprotate35,vprotate40,vprotate45,vprotate55: The camera is tilted 35°/40°/45°/55° relative to the plane and rotates around its normal.

Download the full dataset: world_map.tar.gz (623 MB)

This scene shows a world map fixed to a wall.

It contains the following transformations:

• camrotate0,camrotate22,camrotate45: The camera rotates up to 90° around its view axis, while looking at the posters in a fronto-parallel view or 22°/45° tilted to the side.

• scale: The camera moves away from the scene, resulting in a scaling effect down to 27%.

• vpangle0, vpangle22, vpangle45: The camera starts at a fronto-parallel view, tilting to the side relative to the plane. The optical tilt axis is at 0, 22 and 45 degrees.