SuperPoint TensorRT implementation

[xiang-wuu]

Code in this PR implements baseline PyTorch NN using TensorRT Network API, also provide instructions to run demo script by inferring TensorRT generated engine file along with pre-trained PyTorch weight model, the updated code base of original magicleap/SuperPointPretrainedNetwork is maintained in xiang-wuu/SuperPointPretrainedNetwork

[wang-xinyu]

@xiang-wuu Thanks for the PR!

Can you please share one output image generated from your code?

And does this support tensorrt8.x or 7.x?

Do you have the inference time comparison between pytorch-gpu and your tensorrt implementation?

[xiang-wuu]

@wang-xinyu

Updated the README for sample output result, it's almost identical for PyTorch and TensorRT implementation.

I have tested it only with tensorrt7.x but it should also work with tensorrt8.x as well!

I haven't done the inference time comparison but i put it as separate task under speed/accuracy validation as future work.

[wang-xinyu]

@xiang-wuu Thanks a lot! merging it.

BTW, the 8.x has slightly different APIs from 7.x, please do consider to test with 8.x in the future work if you got any chance.

[xiang-wuu]

@xiang-wuu Thanks a lot! merging it.

BTW, the 8.x has slightly different APIs from 7.x, please do consider to test with 8.x in the future work if you got any chance.

Sure!