Update README.md

README.md

@@ -4,11 +4,25 @@ pipeline_tag: object-detection
 tags:
 - Pose Estimation
 ---
-RTMO Running CUDA DEMO Package
+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.
-- `rtmo_gpu.py`: Defines an RTMO_GPU class, making fine adjustments to CUDA settings.
-- `rtmo_demo.py`: Python main program, which has two arguments:
+- `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.
-- `video`: Contains two test videos provided by Sam's family.
+    - `--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.
+
+Original ONNX models come from [](https://github.com/open-mmlab/mmpose/tree/main/projects/rtmo) trained on body7. We did only 
+
+We did the following to make them work with TensorRTExecutionProvdier 
+
+1. Shape inference
+2. batch size 1,2,4 fixation
+
+Note: TensorrtExecutionProvider only supports Models with fixed batch size (*_batchN.onnx) while CUDAExecutionProvider can run with dynamic batch size.
+
+FP16 ONNX model is also provided.