| --- |
| license: mit |
| pipeline_tag: object-detection |
| tags: |
| - Pose Estimation |
| --- |
| ## RTMO / YOLO-NAS-Pose Inference with CUDAExecutionProvider / TensorrtExecutionProvider DEMO |
|
|
| - `demo.sh`: DEMO main program, which will first install rtmlib, and then use rtmo-s to analyze the .mp4 files in the video folder. |
| - `demo_batch.sh`: Multi-batch version of demo.sh |
| - `rtmo_gpu.py`: Defines an RTMO_GPU (& RTMO_GPU_BATCH) class, making fine adjustments to CUDA & TensorRT settings. |
| - `rtmo_demo.py`: Python main program, which has three arguments: |
| - `path`: The folder location that contains the .mp4 files to be analyzed. |
| - `model_path`: The local path to the ONNX model or a URL pointing to the RTMO model published on mmpose. |
| - `--yolo_nas_pose`: If you run inference with YOLO NAS Pose Model instead of RTMO model. |
| - `rtmo_demo_batch.py`: Multi-batch version of demo_batch.sh |
| - `video`: Contains one test video. |
| |
| # Note |
| |
| * Original ONNX models come from [MMPOSE/RTMO Project Page](https://github.com/open-mmlab/mmpose/tree/main/projects/rtmo) trained on body7. We did only |
| * DEMO Inferecne Code is modified from [rtmlib](https://github.com/Tau-J/rtmlib) |
| * TensorrtExecutionProvider only supports Models with fixed batch size (*_batchN.onnx) while CUDAExecutionProvider can run with dynamic batch size. |
|
|
| We did the following to make them work with TensorRTExecutionProvdier |
|
|
| 1. Shape inference |
| 2. batch size 1,2,4 fixation |
|
|
| PS. FP16 ONNX model is also provided. |
