Probabilistic ToF and Stereo Data Fusion Based on Mixed Pixels Measurement Models
Carlo Dal Mutto, Pietro Zanuttigh, Guido M. Cortelazzo
(Multimedia Technology and Telecommunications Laboratory, University of Padova)
This page contains the supplementary material for the paper "Probabilistic ToF and Stereo Data Fusion Based on Mixed Pixels Measurement Models" by Carlo Dal Mutto, Pietro Zanuttigh and Guido M. Cortelazzo that has been submitted to the IEEE Transaction on Pattern Analysis and Machine Intelligence. Due to the limited space in the paper it is not possible to show all the experimental data and some of the figures in the printed version are too small to appreciate many details of the differences between the results of the various compared methods.
This page contains:
A video showing the behavior of the proposed likelihood functions in different conditions, i.e., smooth areas, depth discontinuities and texture-less areas
The datasets used for the experimental results with the output of the proposed approach and the ground truth data
The code for the proposed site-dependent LBP optimization scheme
Enlarged version of some figures in the paper in order to better appreciate the performances of the proposed method.
The depth discontinuity maps used to compute the results in Table 3 of the paper
Likelihood Functions behavior
Datasets
You can download from here an archive containing all the data used for the experimental results. The archive contains the following data folders:
The folder "Input" contains the input data for the 5 scenes.
The folder "Calibration_Parameters" contains all the calibration parameters for the employed acquisition setup.
The folder "Results" contains the output depth maps computed by the proposed approach on the 5 different scenes used in the experimental results. It has 4 different subfolders with the Tof, Stereo, ML and MAP results (please refer to the paper for the details on how they have been computed). In particualr notice how the ML depth maps still present some artifacts specially in proximity of edges while the proposed MAP optimization leads to sharp and precisely located edges. Note also that the ToF data has a resolution of just 176x144, 16 times smaller (4 times in each direction) than the presented output depth maps and the ToF results in the archive have been interpolated.
The folder "Ground Truth" contains the ground truth data computed with space time stereo.
The folder "Errors" contains the error maps for the ToF, Stereo, ML and MAP results.
The folder "discontinuities" contains the discontinuities maps used for the computation of the results in Table 3.
LBP Optimization
You can download from here an archive containing the LBP optimization code.
ML and MAP comparison
These images shows a comparison of the performance of the ML and MAP schemes (Fig. 15 in the paper). Notice how the edge of the bear's ear is much sharper when the MAP method is employed. Furthermore artifacts like the small region pointing up in the upper edge are removed. Also flying pixels in the middle between the foreground and background are not present in the MAP result while some of them are still present after the ML optimization.
ML
|
MAP
|
Results
These images (Fig. 16 of the paper) contains the output depth maps computed by the proposed approach with the ML and MAP schemes on the 5 different scenes used in the experimental results. The ML depth maps still present some artifacts specially in proximity of edges, while the result of the MAP optimization is much sharper. The ToF data has a resolution of just 176x144, 16 times smaller (4 times in each direction) than the presented output depth maps but still the proposed MAP optimization leads to sharp and precisely located edges.
ML | MAP | |
Scene 1 |
![]() |
![]() |
Scene 2 |
![]() |
![]() |
Scene 3 |
![]() |
![]() |
Scene 4 |
![]() |
![]() |
Scene 5 |
![]() |
![]() |
The discontinuities maps used for the computation of the results in Table 3 are the following (click on the images to view a full-resolution version):
Scene 1 | Scene 2 | Scene 3 | Scene 4 | Scene 5 |
![]() |
![]() |
![]() |
![]() |
![]() |
If you are interested in our research you can visit our website: http://lttm.dei.unipd.it
For any question or clarifications about this datasets please write to: zanuttigh@dei.unipd.it