Cool work. Which one is faster, FDSCS or RealtimeStereo on TX2.

[leejiajun]

The work, FDSCS, also provides a low GPOS model, please see https://github.com/ayanc/fdscs, while your paper doesn't mention it.

[Abhishekvats1997]

@leejiajun were you able to train this ?

[leejiajun]

@Abhishekvats1997 You mean FDSCS? Yes, I was.

[Abhishekvats1997]

@leejiajun I meant RealtimeStereo the current repo

[leejiajun]

@Abhishekvats1997 Seems the run.sh uses the previous architecture. The architecture--stackhourglass--is not able to run, so I guess you need to change the architecture to RTStereoNet.

[Abhishekvats1997]

@leejiajun yup just saw that Also if i may ask, were you able to find any work which is strikes a good balance between accuracy and speed ? I am looking to deploy stereo based models on Nvidia Jetson. So far my best hope is AnyNet.

[leejiajun]

@Abhishekvats1997 I deployed FDSCS on a NPU. I think RTStereoNet and AnyNet are both good choices. Also, according to the paper of RTStereoNet, this net has lower FLOPs and lower PC error than AnyNet. If you desire accelerating further on Jetson, I guess you would like to quantize these models.

[Abhishekvats1997]

@leejiajun i tried this one against the AnyNet on real world samples that i collected. AnyNet seems to perform much better. Did you have a look at DeepPruner Fast ? It has amazing results even better than PSM Net and much faster.

[leejiajun]

@Abhishekvats1997 I couldn't search any paper's title called DeepPruner Fast. As far as I know, FDSCS is faster than DeepPruner.

[leejiajun]

I see. DeepPruner Fast is a variant of DeepPruner. 😂