MULTIVISION

Esta web ha sido optimizada para exploradores compatibles con estándares - Firefox, Chrome, Safari...

  • Aumentar fuente
  • Fuente predeterminada
  • Disminuir fuente

Downloads

This website provides a dataset with ground truth segmentation which enables the objective evaluation of background subtraction algorithms based on depth information as well as color images. The dataset consists of two different sets of sequences: the first set has been recorded by using stereo cameras combined with three different disparity estimation algorithms; the second one has been recorded by using the Kinect sensor from Microsoft.

Background subtraction model based on color and depth cues (Stereo disparity depth information)

If you use these sequences to test and report results in any publication, we request you to kindly acknowledge this website (http://atcproyectos.ugr.es/mvision/) and cite the following paper that summarizes this part of the dataset, performance evaluation metrics, and the results and comparison with previous methods:

  • Fernandez-Sanchez E.J., Rubio L., Diaz J., Ros E. Background subtraction model based on color and depth cues. Machine Vision and Applications, Journal of. ISSN: 0932-8092. pp:1-15. Springer Link ; BibTex

In this work we have provided a dataset for evaluation of background subtraction techniques using depth-computation algorithms. In order to enable this, four sequences have been recorded with rectified stereo cameras, and some frames have been hand-segmented to provide ground-truth information.

In addition, disparity information obtained by three different approaches is also available.

 

Suitcase

In this sequence, a person enters the scene with a suitcase and leaves it on the floor. The main difficulty of the sequence is the low lighting and color saturation, as well as the similar color between the suitcase and the floor.

Sample images of "Suitcase" sequence

Links:

Original stereo images. (757 MB)

Ground truth. (5.4 MB)

Phase-based disparity. (108 MB)

Variational-based disparity. (11 MB)

SGBM-based disparity. (48 MB)

 

 

Crossing

Two people walk in and out of the camera field. The dark floor complicates the detection by color when they get near the camera, while the range is less useful when they get near the wall.

Sample images from "Crossing" sequence

 

Links:

Original stereo images. (1.5 GB)

Ground truth. (7.9 MB)

Phase-based disparity. (26 MB)

Variational-based disparity. (24 MB)

SGBM-based disparity. (103 MB)

 

 

LCDScreen

A person walks into a lab and deposits a black box in front of a black LCD screen. In addition, there are flickering lights in the ceiling.

Sample images from "LCDScreen" sequence

Links:

Original stereo images. (384 MB)

Ground truth. (1.8 MB)

Phase-based disparity. (6.5 MB)

Variational-based disparity. (9.5 MB)

SGBM-based disparity. (11 MB)

 

 

LabDoor

A person walks in and out of the camera field, projecting shadows on background objects. In addition, there are occlusions due to background objects, flickering lights in the ceiling and sudden illumination changes.

 

 

Links:

 

Original stereo images. (1.7 GB)

Ground truth. (4 MB)

Phase-based disparity. (45 MB)

Variational-based disparity. (38 MB)

SGBM-based disparity. (65 MB)

 

 



 

 

Background subtraction based on color and depth using active sensors (Kinect depth information)

If you use these sequences to test and report results in any publication, we request you to kindly acknowledge this website (http://atcproyectos.ugr.es/mvision/) and cite the following paper that summarizes the dataset, performance evaluation metrics, and the results and comparison with previous methods:

  • Fernandez-Sanchez E.J., Diaz J., Ros E. Background Subtraction Based on Color and Depth Using Active Sensors. Sensors 2013, 13, 8895-8915. Sensors ; Bibtex

In this work we have provided a dataset for evaluation of background subtraction techniques using the Microsoft Kinect sensor. In order to enable this, we have recorded four sequences (both RGB images and depth), and some frames have been hand-segmented to provide ground-truth information.

 

ChairBox

In this sequence, a person enters the scene with a box and leaves it on a chair. There are flickering lights as well as areas where depth cannot be obtained by the Kinect.

Links:

Images, Depth and Ground Truth. (332 MB)

 

 

Wall

In this sequence, a flat object appears close to a wall, creating shadows and highlighted regions. The main difficulties are the similarity of depth values between foreground and background and the changes of lighting.

Links:

Images, Depth and Ground Truth. (134 MB)

 

 

Shelves

In this sequence, a person enters the scene and puts two objects on shelves. There are changes of exposure as well as difficult depth estimation.

Links:

Images, Depth and Ground Truth. (409 MB)

 

 

Hallway

This sequence has been recorded aiming to a hallway. There are reflections, complicated lighting, objects with colors similar to the background, and sudden illumination changes.

Links:

Images, Depth and Ground Truth. (419 MB)