train monoResmatch

[manuel-88]

Hi, thanks for uploading the code. The training should work without proxy labels in a self-supervised manner? Meaning I should be able to run the training without the --data_path_proxy [path_cityscapes_proxy] argument. BR

[fabiotosi92]

Hi, in the current version of monoResMatch you need proxy labels in order to train the network. If you don't have them, you should edit the code accordingly.

[manuel-88]

Are this proxy labels somewhere available to download?

[fabiotosi92]

Since proxy labels require a large amount of space, they are not available to download. However, you can use this code (https://github.com/ivankreso/stereo-vision/tree/master/reconstruction/base/rSGM) to generate them as 16 bit images in a very short time.

[manuel-88]

Thank you for the information

[manuel-88]

I run the SGM code you suggested. The output of the tool is a 8 Bit image. How do I convert to 16 Bit and what is the advantage?

[fabiotosi92]

Hi, we use 16 bit images because proxy labels (or, in general, grountruth depth data) could encode sub-pixel information. You can save disparities using OpenCV as follows:

include <opencv2/opencv.hpp>

...

Mat disp = Mat(img.getHeight(), img.getWidth(), CV_8UC1, (uchar)img.getData()); disp.convertTo(disp, CV_16UC1); disp=256.0; imwrite(output_img, disp);

[manuel-88]

Thanks for the code. But do you not lose the subpixel accuracy when casting to 8 Bit? And why do you multiply the values with 256? Then all disparities which are greater then 256 reach the 16 Bit limit.

[fabiotosi92]

The rSGM code does not provide sub-pixel information, but it could be easily extended for the purpose (e.g., implementing parabola interpolation). However, in accordance with the KITTI dataset, we mantain the standard disparity values range [0..256] encoding such information as 16 bit PNG images.

[manuel-88]

ok, I understand now. Thank you.