Best way to implement custom models that are not Detection or Segmentation

[Red-Draken]

Description

So I am trying to implement ultra fast lane detection v2 in my deepstream application. The ‘nvinfer’ plugin only supports detection and segmentation models so I cannot use that for this purpose. The steps I can think of would involve,

• A probe to get the relevant frame and associated metadata
• Do inference on the frame using TensorRT optimized engine of ultra fast lane detection v2
• Attach the inference data to NvDsBatchMeta using NvDsUserMeta
• Use another probe to set displaymeta to visualize lanes

Now, the issue is, this sounds very over the top and I am hoping there is a better way. Am I missing something? any help would be appreciated.

Environment

TensorRT Version: 8.4.15
GPU Type: RTX 3080
CUDA Version: 11.7
CUDNN Version: 8.7.0
Operating System + Version: Ubuntu 20.04
Deepstream Version: 6.1.1

[spolisetty]

@Red-Draken

So I am trying to implement ultra fast lane detection v2 in my deepstream application.

Hi,

This looks more related to Deepstream, we are moving this post to the Deepstream forum to get better help.

Thank you.

[Fiona.Chen]

You are in the right way, the model needs specific post-processing and the output is special too, you need the customization mentioned by you.

[Red-Draken]

Thank you for prompt response Fiona. I will open a new question in case something comes up but I am good for now.