beaupreda/semi-global-matching

Semi-Global Matching

Semi-Global Matching

Implementation of the Semi-Global Matching algorithm in Python.

Dependencies

• OpenCV
• NumPy

Instructions

$ git clone https://github.com/beaupreda/semi-global-matching.git
$ cd path/to/semi-global-matching

Usage

python3 sgm.py --left [LEFT IMAGE NAME] --right [RIGHT IMAGE NAME] --left_gt [LEFT GT IMAGE NAME] --right_gt [RIGHT GT IMAGE NAME] --output [OUTPUT IMAGE NAME] --disp [MAXIMUM DISPARITY] --images [TRUE OR FALSE] --eval [TRUE OR FALSE]

Example

python3 sgm.py --left cones/im2.png --right cones/im6.png --left_gt cones/disp2.png --right_gt cones/disp6.png --output disparity_map.png --disp 64 --images False --eval True

Other implementations

• C++
• MATLAB
• CUDA

References

• Stereo Processing by Semi-Global Matching and Mutual Information
• LUNOKHOD SGM Blog Post

semi-global-matching's Issues

OverflowError: Python int too large to convert to C long

When using your example:
python sgm.py --left cones/im2.png --right cones/im6.png --left_gt cones/disp2.png --right_gt cones/disp6.png --output disparity_map.png --disp 64 --images False --eval True
I get the following logs:

Loading images...
Starting cost computation...
Computing left and right census...Traceback (most recent call last):
File "sgm.py", line 431, in
sgm()
File "sgm.py", line 392, in sgm
left_cost_volume, right_cost_volume = compute_costs(left, right, parameters, save_images)
File "sgm.py", line 261, in compute_costs
left_img_census[y, x] = np.uint8(left_census)
OverflowError: Python int too large to convert to C long

please delete

My bad, please delete the issue.

Recall function is wrong

The in the following lines of the get_recall method,

gt = np.int16(gt / np.amax(gt) * float(args.disp))
disparity = np.int16(np.float32(disparity) / np.amax(disparity) * float(args.disp))

the np.amax(...) functions should be replaced with 255.0.

Using the maximum disparity found makes no sense because it influences the value of all the other disparities.

Does this algorithm work on unrectified images?

Hi,

I was wondering if it was possible to use this algorithm on unrectified images.

Thank you!

Can't find the output image

Hey, I've tried your code with the example command. And I can't find this output image disparity_map.png, but left_disparity_map.png and right_disparity_map.png.
Besides, I wanna ask, are these two pictures under cones folder (disp2.png and disp6.png) generated before the command is executed? Because I changed the command python3 sgm.py --left cones/im2.png --right cones/im6.png --left_gt cones/dis2.png --right_gt cones/dis6.png --output disparity_map.png --disp 64 --images False --eval True and could also be executed successfully without the output image.

About post_processing

Thank you for sharing the code!
I test it on middlebury 2006 and the recall rate is about 10%, which is too low to use.
I found there is no consistency check and small patches are now removed, but I've learned they won't improve that much.
I wondered if there are some other possible explanations about this question.
Thank you!

Why the processing time is so long?

First thanks for provide the code.
I use the code to generate disparity map with with KITTI RGB stereo images, and the processing time is nearly 1800 seconds for only one pair images.
The result of stereo matching is good, but why the processing is so long? Is anywhere wrong?

How to change the cost function

I need to test 2 different descriptors (HOG and BRIEF) for my computer vision class and I'm trying to find where they could be used in the code. I believe it would be in the compute_costs function, where the census transform is used (this seems to be very similar to BRIEF).

Am I correct?

Finding Disparity Range

It's taking me quite a long time to look for the disparity range value to set in the command line to get a smooth disparity output from the code. Is there a good way to figure out what disparity range to set the "--disp" parameter to?