search

siposcsaba89 / stereo-sgm-opencl

SGM implementation
52 stars 16 forks source link

sgm-opencl implementation with costs cube as argument #8

Open squirelt opened 1 year ago

squirelt commented 1 year ago

Hi again :)

I'm struggling a little bit to understand the code precisely. The reason I'd like to understand everything is that I want to adjust it to my needs. More precisely, I'd like to hand over a cost cube [M,N,D] to the aggregation instead of left and right census transform. This makes the code more flexible to me since I have to test a wide range of cost functions.

Is there any documentation for the code? Or did you do something similar?

Thanks for your help!

siposcsaba89 commented 1 year ago

Hi,

This implementation is a port from fixstars's libsgm implemented in cuda to opencl (https://github.com/fixstars/libSGM) like 2 years ago version . They continuously improve the performance of their library so maybe it is worth to check out. To tell you the truth it was like 2 years ago, when I did this, so I forget a lot.

This article is a good starting point to understand what is happening inside the individual kernels, maybe you already know it, but I link it anyway: https://arxiv.org/pdf/1610.04121.pdf

If I will have more time, I will take a look at those kernels to see if I can help you more.

squirelt commented 1 year ago

Hi, Eventually I managed to adapt the software such as it takes a cost cube now as input. As mentioned, this gives me the flexibility to use diverse cost functions. If you are interested I'll share it with you. (In the end after really understanding what the software does, the adaption was quite straight forward).

And thanks again for your support!

siposcsaba89 commented 1 year ago

Hi, Thanks for letting me know, great work!