Why stereo.cc is not considered to be Cuda enhanced ?

[AhmedElsafy]

I see in stereo.cc it includes system.h same as mono.cc, which calls the ORBExtractor.cc which is cuda enhanced. I don't understand why it is considered not to be cuda optimized since it calls the same system object ?

[cyclojc]

I have the same question

[thien94]

Hi @AhmedElsafy and @cyclojc , thank you for the question and sorry for the late reply.

To clarify: the CUDA enhanced components also apply to the stereo node. The "not supported" description is about testing and verifying that the nodes are indeed working as intended. Meaning, I haven't done any verification that the new stereo/RGB-D node are enhanced compared to the original ORB-SLAM2.

You can still be able to run them without any problems.

[zzjkf2009]

Based on my test result (RGBD version) running on Jetson NX, it seems like that GPUs are used, but the tracking performance is basically the same as the one without GPU enhanced. Any suggestion?

[cyclojc]

@zzjkf2009 I had the same result, same question (using stereo version)

[thien94]

@zzjkf2009 @cyclojc for RGB-D, there are a number of newer packages that have addressed the performance and other issues that I would suggest checking out:

• Dataflow ORB-SLAM2: code, paper,
• gf-orb-slam2: code,
• ORB-SLAM2-GPU-RGBD: code.

As for the performance of this repo, you can first replicate the same experiment with the original ORB-SLAM2 and see if there are any differences. If the problems are the same, then it would be the limitation of the underlying method. The advantage of using CUDA, at least for this repo, is to accelerate certain bottlenecks of the system. Besides that, nothing should be different performance-wise.