From mmdetection to DeepStream-Yolo: A success story

[jstumpin]

Hi folks, just a quick update on importing a fine-tuned RTMDet model from mmdetection to this repo. Here how it goes:

1. Fine-tuning
   Install mmdetection on RTX 50xx-series:
   https://gitee.com/Wilson_Lws/MuseTalk-50Series-Adaptation/blob/master/README.md
   Import/export dataset from Ultralytics to COCO:
   https://docs.voxel51.com/user_guide/import_datasets.html#yolov5
   https://docs.voxel51.com/user_guide/export_datasets.html#coco
   Train using custom dataset:
   https://github.com/open-mmlab/mmdetection/blob/main/demo/MMDet_Tutorial.ipynb
   Visualize using SwanLab (optional):
   https://docs.swanlab.cn/en/guide_cloud/integration/integration-mmdetection.html

2. ONNX conversion
   Install mmyolo on RTX 40xx-series:
   https://github.com/marcoslucianops/DeepStream-Yolo/blob/master/docs/RTMDet.md#1-download-the-rtmdet-mmyolo-repo-and-install-the-requirements (ensure to use the same Torch/CUDA versions as the ones used in 1.)
   Convert fine-tuned mmdetection model to mmyolo model:
   https://github.com/open-mmlab/mmyolo/blob/main/tools/model_converters/rtmdet_to_mmyolo.py
   Convert model to ONNX:
   https://github.com/marcoslucianops/DeepStream-Yolo/blob/master/docs/RTMDet.md#4-convert-model

The recipe was only tested using these configurations:
https://github.com/marcoslucianops/DeepStream-Yolo?tab=readme-ov-file#deepstream-71-on-x86-platform