FBK home > INFORMATION TECHNOLOGY > Technologies of Vision > Databases

Technologies of Vision: labeled databases sharing

DIPLODOC - A labelled road sequence taken from an in-vehicle stereo camera

Notice: Available for download. This stereo image archive is provided for research or academic purposes only. Publications that include results obtained with this database should please refer to DIPLODOC road stereo sequence FBK Technical Report Nr 164010, M.Zanin, S. Messelodi, C.M. Modena

* A sequence of 865 image pairs taken from a stereo camera mounted on a moving vehicle is available for download. You can use these images to benchmark you stereo algorithm for road and obstacle detection.
The road region is manually segmented and it is proposed as ground truth for performance evaluation of road detection algorithms.
The sequence is the composition of five subsequences, each of them presenting very different traffic and road conditions: from highway-like roads to urban scenarios with crossroads, parking lots and complex environment; from congested traffic to completely free road.
The images were captured on July, 16 2004, about 11 a.m., near Trento, ITALY. The following map shows where the subsequences were taken.

Map of the sequence location
Click on it to view a full resolution map (1617x1023 550KB)

* The acquisition device is a Videre Design MEGA-D stereo camera pair installed near the rearview mirror. The sequence is 15 fps, 320x240, color. The images are saved in a lossless format (PNG) without any pre-processing.
The stereo pair is calibrated with the SVS software and the camera parameters are included in the downloadable file. In the following figure there is an example of left and right images from the sequence (frame no. 202):

example of image pair

* The definition of road given to the operator was: "everywhere a car could drive without going up a step". The ground truth for an image is saved as a set of polygons. Some of them representing ideal road regions, some of them representing objects that occlude portions of the road. From these regions it is possible to compute the road region actually visible in the considered image. The following figure presents an example (frame no. 202), with one road region and two occluding regions. The last image is the derived visible road region.

example of road and occlusion polygons
example of visible road region


* The complete 865 frames long stereo sequence, with ground truth data and camera parameters is available for downloading in a 214MB file. In order to download it, please fill in the following form. An email with a link to the file will be sent to the submitted email address.

Please submit at least your email address.
In respect of your privacy, your email account information will only be used to communicate the link to the labelled data. Responsible of the e-mail address collection is Michele Zanin (FBK). From 01.01.2013 the responsible is Carla Maria Modena. For further information on the data or the project, please contact Stefano Messelodi: messelod (at) fbk (dot) eu
your name
e mail (required)
web site


DIPLODOC project, including the collection of this database, was supported by PAT, Provincia Autonoma di Trento, Fondo Unico per la Ricerca.


For further information on the data or on the DIPLODOC project, please contact Stefano Messelodi, e-mail messelod (at) fbk. eu

This work is licensed under a Creative Commons License Creative Commons License