Investigate lidar_apollo_segmentation_tvm detection quality

[ambroise-arm]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.
• I've agreed with the maintainers that I can plan this task.

Description

I have tested the TVM package replacing the TRT package in the rosbag demo as part of the PR: #1181 (comment). And the quality of the detection wasn't as good as the detection from the TRT package, at least visually. Some cars were not detected, the lines and circles which indicate the velocity (or planned trajectory, I'm not sure) were wrong, at least on parked vehicles. And the type of detected vehicles might be wrong, it would need to be checked.

Purpose

Get the detection quality of the TVM package on par with the TRT package.

Possible approaches

There are multiple possible causes for the difference:

• The TVM version of the package is derived from an old version from TierIV; I expect the TRT version in this repository is more recent
• The neural network may have been updated since then; trained differently/better
• The port may have introduced issues/bugs
• The parameters may not be correct or not all there

Definition of done

• confirm and document the issue
• find the cause(s)
• update the lidar_apollo_segmentation_tvm* packages and/or the neural network in use, depending on what is necessary

[angry-crab]

#1880 has to be merge first.

[angry-crab]

Issue confirmed. There are quite a lot of false positive detections. This is the output of detected_object_feature_remover. And it seems that all of the objects are removed by other filters.

[angry-crab]

According to @yukkysaito san, the weights have been update.
@ambroise-arm I see that the current pipeline only supports onnx files. However, lidar_apollo_segmentation is using caffe model. Are we planning to support other types of model in the future? Or shall we convert it to onnx format?

[angry-crab]

Converting to onnx format seems better since caffe operations are not fully supported and will probably be removed in the near future.
https://discuss.tvm.apache.org/t/caffe2-frontend-support-is-being-dropped-to-unblock-pytorch-update/12442

[ambroise-arm]

@angry-crab I don't think the Autoware pipeline for tvm only supports onnx, and other formats could be used. It is really down to what TVM supports. If caffe support is dropped in TVM, then I agree that it can't be used as is and needs to be converted.

[angry-crab]

To convert caffe models into onnx, this might be a good choise.
https://hub.docker.com/_/microsoft-windowscamerateam-onnxconverter

[ambroise-arm]

I couldn't convert the current caffemodel files with the microsoft tool, it complains about the "Slice" layer.
I've also tried to use small tools like https://github.com/asiryan/caffe2onnx, which clains to support "Slice", but there is a "Silence" layer in the network that isn't supported.

[angry-crab]

I couldn't convert the current caffemodel files with the microsoft tool, it complains about the "Slice" layer. I've also tried to use small tools like https://github.com/asiryan/caffe2onnx, which clains to support "Slice", but there is a "Silence" layer in the network that isn't supported.

I got it work using the converted model. It looks good. caffemodel file contains weights while prototxt file describes the structure of network, which does not have Slice and Silence layers.