Add polyp detection application

[yiheng-wang-nv]

No description provided.

[jjomier]

@yiheng-wang-nv can you look at the linting issues?
Let us know when this is ready to review.

[yiheng-wang-nv]

Thanks @jjomier , the lint issue is fixed. I will update the draft PR for adding more useful operators, and then convert to ready for review

[yiheng-wang-nv]

Hi @jjomier , could you please help to review this PR? Thanks in advance.

[jjomier]

Thanks @yiheng-wang-nv.
Since we are using an ONNX model, why aren't you using the Holoscan ONNX inference operator instead of converting to TRT (that operator does the conversion automatically).? Ideally if we can using the base Holohub/Holoscan docker image that would be ideal. Why do we need Torch?

There are still some linting errors. You can run do a ./run lint if you want to check for the errors before commiting.

[tbirdso]

Hi @yiheng-wang-nv could you please remove the [draft] label if this PR is ready for review+merge?

[yiheng-wang-nv]

Thanks @yiheng-wang-nv. Since we are using an ONNX model, why aren't you using the Holoscan ONNX inference operator instead of converting to TRT (that operator does the conversion automatically).? Ideally if we can using the base Holohub/Holoscan docker image that would be ideal. Why do we need Torch?

There are still some linting errors. You can run do a ./run lint if you want to check for the errors before commiting.

Thanks @jjomier for the suggestions!

• I will check how to use the onnx inference operator.
• As for the Holoscan docker image, do you have any references? I took the sam2 application for reference, thus used the pytorch base image.
• As for the linting errors, I will fix them.

[Nic-Ma]

Hi @yiheng-wang-nv ,

Let's try to complete this PR and merge it.

Thanks.

[jjomier]

@yiheng-wang-nv, I believe we are over-complicating this submission. Why don't we use the current polyp segmentation application but add the new ONNX model instead? Is the ONNX model publicly open or has it been fined tuned?

[yiheng-wang-nv]

current polyp segmentation application

Hi @jjomier , thanks for the suggestion. For this PR, the model I used is trained by myself with the real colon dataset, and the model is used for detection instead of segmentation. Therefore, we may not be able to use the existing model instead.

[yiheng-wang-nv]

TODO: the URL need to be replaced by a NGC catalog link

[Nic-Ma]

Hi @yiheng-wang-nv , @jjomier ,

May I know the remaining tasks of this PR?

Thanks.

[jjomier]

Hi @yiheng-wang-nv , @jjomier ,

May I know the remaining tasks of this PR?

Thanks.

Working on releasing the model first