How to use this page
- Fill in the appropriate data in all the form-fields of the previous page.
- After submitting your job, our server will try to fetch your images,
convert them into an internal floating-point image format, do some
preliminary checks and finally start the calculations.
- You will be shown a log-file of your request, where you can watch this
process. The results will simply be attached to the end of the log-file.
- If your browser can not handle automatic updates (most browsers can),
you have to push the RELOAD-button of your browser once in a while to get
the current status.
- After your calculations are finished, the results are stored at our
server for further reference. However, after one or two days, the data
will no longer be available to you.
When things go wrong
- Main URL:
This points to the location of your data.
Typically, stereo images are stored as one single chunk of data, but in
various different formats. Choose the appropriate format for conversion. The
table below list all available formats:
This is the standard arrangment. You have to supply two URLs;
one for the left image, one (in the secondary input-field) for the
The images are arranged for parallel viewing; this means the left stereo image
is left, the right stereo image left.
The images are arranged for crossed-eye viewing; thus the left
stereo image is right, the right stereo image is left.
Some stereo images are arranaged for viewing with a small mirror;
one of the images seems to be the mirror image of the other
(actually, it's not - otherwise you wouldn't see 3D...)
This one is (obviously) for anaglyph stereo images. In these images,
the left data is found in the red color channel, the right stereo data
in the green or blue channel.
Some stereo images are arranged in this order, to allow
simultaniously parallel and crossed-eye viewing.
Here, the stereo images are stacked vertically above each other. The
right image is on top, the left image on the bottom of the stereo
The left stereo image is vertically stacked above the right stereo
Both stereo images are interlaced; even horizontal image rows
correspond to the left, odd horizontal image rows to the right
Image is assumed to be a Single Image Random-Dot
Stereogram. The algorithm tries the see the
3D-effect in the SIRD, but note that the implementation might encounter
the same difficulties as you by viewing such images, i.e., not
seeing any depth ...
If you specify "Two Images", you have to fill in the secondary URL with the
pointer to the right image data.
Normally, horizontal disparities are calculated. However, if your
images have vertical disparities, click on the appropriate button.
- Maximal Image Size:
The images are zoomed down to this
image-size before starting processing.
The implemented network has a specific fusion range,
similar to human stereo vision. Before disparity calculations, your stereo
images have to be aligned accordingly.
You can select either:
- Automatic Vergence Control: the algorithm tries
to align the stereo images automatically. This might fail, than
try "Fixed Disparity Range" instead.
- Total disparity range:
this sets the fusion range of the network.
The fusion range of the network is restricted to maximal
- Allow for rotation:
the vergence system includes a cyclovergence unit. In some cases,
you might get better results if you disable that. The vergence
system uses then only horizontal or vertical panning for
specifies the area of the image which is considered "important" by
the vergence system. For images with mainly frontoparallel surfaces,
try the "wide"-setting;
for small objects in the center of the stereo images, try "narrow".
- Fixed Disparity Range: if you know that your images
have no vertical disparity component, and if you know the range
of horizontal disparities in advance, this is the setting for you.
- Spatial Scales used:
Like in human stereo vision, data is
processed in parallel at several spatial scales. Since this is a serial
machine, selecting less spatial scales speeds up processing. Using more
spatial scales gives smoother results, but the improvement is not dramatic.
- Disparity Detector:
The different units
implement various methods for disparity estimation. Try any one of these!
Gradient-based: Uses standard optical flow estimation
Tensor-based: Based on an algorithm for the estimation of
texture orientation. It searches for an axis through the origin of
the Fourier power spectrum with minimal inertia.
Gradient-based, iterative: iteratively refines intial
disparity-estimate. Is very precise, but is not really needed with
coherence-based stereo. In addition, it runs slow.
Robust Gradient: Basically identical to the gradient-based detector,
except the estimation stage; here, a robust estimator is used instead of the
usual least-square estimator.
Complex-Cell: calculates motion-energy for various
disparities, and selects the one with strongest response. Might also be
realized by neural circuitry in human visual cortex.
Windowed Fourier-Phase: utilizes the Fourier-Shift-Theorem
to calculate the disparity out of local phase-shifts; very simple
The coherence stage is the most important stage of the new algorithm. Since
it's a parallel algorithm implemented on a normal computer, two different
implementations are available:
Fast Sync-Scheme: Data from the coarser spatial channels is
simply included in the disparity stack of the finest resolution layer, where
the coherence-detection is performed. This is simple and fast.
Hierarchical Sync-Scheme: This processing option is more adaptive -
it takes into account that disparity data from coarser spatial channels is
less precise than from finer spatial channels (due the spatial averaging in
the coarse spatial channels). Gives better results, but takes longer. If
spatial coherence is also included, the results improve a little bit
- but the algorithm runs longer.
Verify Data: If set to a value larger than 0.0, the disparity map is
set to zero wherever the validation map is below that threshold. Use values
between 0.0 and 1.0 units.
Your can choose between different output-formats:
Colored Disparity Map : with this output format, the disparity map is
converted into a shaded color image, where different colors decoded depth.
Takes long, but can display very fine depth-variations.
Grayscale Display : this output format gives back the various
calculated maps without any conversion applied.
Red_blue Anaglyph : your stereo image is converted into an anaglyph
which may be viewed with a red filter in front of the left eye, and a blue
(or green) filter in front of the right eye.
- Most problems occur in accessing your data by our server. We can
handle http-, ftp- and gopher-protocol. Be sure to
give a correct URL. You might
check the URL you supplied to our software by opening this location in your
own browser. But be sure to check by opening over the WEB, not
locally. If your browser can fetch your data, the URL should be fine. If our
server still can't fetch your data, please contact us.
- The algorithms have been designed to be fairly bullet-proof, but if you
do not enter correct values, you won't get any sensible results.
- This server is used also other purposes. If a certain load-average is
exceeded, this service is temporarily suspended.
© 1994-2000 - all rights reserved.