Stereo and Depth-from-Defocus Dataset 2: The average office


The average office: The original POV-Ray scene

From the original page:

Unbelievable! I've finished this scene! It was at first a typical office made from remembrances of offices I worked at, but later evolved onto a real office, as a friend of mine wanted to test how things would look on his new property (see WIPs 1 & 2).

Now it is again a ficticious office, wich I populated with some more objects done with Wings3D.

Copyright 2012 Jaime Vives Piqueres (jaime-at-ignorancia-dot-org)

Stereo and Depth-from-Defocus dataset


Here are JPEG-encoded previews for all the images in the dataset. The dataset includes better quality images, encoded as 16-bits 1080x720 PNG. Click on low-resolution images to get a high-resolution preview.

Note that there are small low-frequency illumination differences between the left and right images, which are due to radiosity computation (if someone knows how to share the radiosity computations between several views, please tell me).

All-in focus

office_stereo1_all office_stereo2_all

Far focus, large and small depth-of-focus

office_stereo1_far office_stereo2_far office_stereo1_far2 office_stereo2_far2

Near focus, large and small depth-of-focus

office_stereo1_near office_stereo2_near office_stereo1_near2 office_stereo2_near2

Depth, disparity and occlusion map ground truth

Important note: the smoke of the cigarette put on the nearest desk generates wrong depth/disparity values

office_stereo12 office_stereo21 office_stereo12.occ office_stereo21.occ

Stereo and focal blur parameters

The interocular is 8 (see variable half_interocular in the pov files), and the convergence (or zero-disparity) distance (see variable convergence_dist in the pov files) is 500.

The *_near images are focused on (65,289,-165) (disparity=-11.7), the *_far images are focused on (-100,110,20) (disparity=0.9), and the *_all images are all-in-focus.

The images with a small depth-of-focus were rendered with an aperture setting of 10, which means that the iris is a disc with diameter 40 (5 times the interocular) in world units.

The images with a larger depth-of-focus were rendered with an aperture setting of 5, which means that the iris is a disc with diameter 20 (2.5 times the interocular) in world units.


If you use this dataset, please cite the original publication where it appeared:

F. Devernay, S. Pujades and V. Ch.A.V., ``Focus mismatch detection in stereoscopic content'', Stereoscopic Displays & Applications XXIII, Woods, Holliman, Favalora (eds.), Proc. SPIE 8288, paper 8288-12, 2012 (HAL, presentation video).

We are also preparing a journal paper which extends the original paper with new interesting results, and hope it will be published in 2013. Please check out my publication list to see if it's already there.


Re-rendering the dataset

Here are the instructions if you want to re-render the images from the original POV-Ray files. This way, you can adjust the camera parameters to your needs.

POV-Ray files

The files can be rendered using MegaPOV 1.2.1 with the annotation patch, and POV-Ray 3.7 RC6 (POV-Ray 3.6 generates visible artifacts between the left and right views). On Mac OS X, MegaPOV has to be compiled with GCC 4.2 (not LLVM-GCC or clang), or the rendering will be incorrect (I have instructions to install GCC 4.2 on recent Mac OS X).


If no radiosity is applied, the rendering will be completely wrong: the shadowed aread (under the arcades) will be very dark.

In order to apply radiosity:

Change the parameteres in the "control center" section of office/office.pov for 1st pass.

then, run one rendering. This will save a radiosity file "office_rad_custom" in the current directory.

For subsequent renderings from the same point of view (it is a good idea to render the left and right views in separate directories), change the parameters for final pass.

This will load the radiosity file and thus save computations.

All-in-focus view are rendered without focal blur, but with antialiasing (thus the options +A0.0 +J0.0 for the office_stereo*_all.png images, which mean to perform 3x3 antialiasing on every pixel). The +FN16 option means to render 16-bits PNG images.



Left view:

Right view (+K0.0 is changed to +K1.0):

Ground-truth disparity and occlusion maps

The ground truth disparity map can be obtained using the VLPovUtils utilities. Here are the commands to produce this data, once the office_stereo1_all and office_stereo2_all scene have been rendered.

The first command computes the motion field and occlusion maps. The results are stored in office-left/, office-left/office_stereo1_megapov.office_stereo2_megapov.occ.tif, office-right/, office-right/, office-right/office_stereo2_megapov.office_stereo1_megapov.occ.tif. The disparity maps are in the *.mx.tif files (stored as 32-bit floating point TIFF images) and the occlusion maps in the *.occ.tif files (stored as 8-bit grayscale TIFF files, converting them to PNG saves a lot of disk space).

The stereo disparity corresponding to the two focus distances can be obtained using the vlpov_project utility. The focal points are (-100,110,20) [far] and (65,289,-165) [near] in POV-Ray coordinates. Note the Z coordinate of the 3D points has to be reversed, due to the left-handedness of POV-Ray's reference frame (see the code of function vlpov_get_cam in vlpovutils). The pixel coordinates of each points are in the first two columns, and the disparities (0.9 and -11.7) are in the fourth column of the output (last column, the vertical disparity, is almost zero):

Valid XHTML 1.1! Valid CSS! Level Double-A conformance icon, W3C-WAI Web Content Accessibility Guidelines 1.0