trt_pose_demo

A Demo Application for NVIDIA TensorRT Pose Estimation

What does this application do?

This application just displays pose estimation results for camera-captured video or H.264/H.265 video file with NVIDIA TensorRT pose estimation. This application also has a capability to save the results in a CSV file.

Prerequisites

• NVIDIA Jetson Nano/TX2/AGX Xavier Developer Kit with JetPack installed
• USB web camera or MIPI CSI camera like Raspberry Pi v2 camera

Setup

1. Install PyTorch and torchvision. Please refer to this page.
2. Install NVIDIA-AI-IOT/torch2trt.
3. Install NVIDIA-AI-IOT/trt_pose.
4. Install this application.

   $ git clone https://github.com/MACNICA-CLAVIS-NV/trt_pose_demo
   $ cd trt_pose_demo
   

5. Copy the resnet18_baseline_att_224x224_A model file to the application directory.
6. Copy the human_pose.json task description file to the application directory.

Usage

First, clock up your Jetson.

$ sudo nvpmodel -m 0
$ sudo jetson_clocks

The following command starts this application.

$ python3 trt_pose_app.py [-h] [--camera CAMERA_NUM] [--width WIDTH]
                       [--height HEIGHT] [--fps FPS] [--qsize QSIZE] [--qinfo]
                       [--mjpg] [--title TITLE] [--nodrop] [--repeat] [--h265]
                       [--model MODEL] [--task TASK_DESC] [--csv MAX_CSV_REC]
                       [--csvpath CSV_PATH] [--verbose]
                       [SRC_FILE]

TRT Pose Demo

positional arguments:
  SRC_FILE              Source video file

optional arguments:
  -h, --help            show this help message and exit
  --camera CAMERA_NUM, -c CAMERA_NUM
                        Camera number, use any negative integer for MIPI-CSI
  --width WIDTH         Capture width
  --height HEIGHT       Capture height
  --fps FPS             Capture frame rate
  --qsize QSIZE         Capture queue size
  --qinfo               If set, print queue status information
  --mjpg                If set, capture video in motion jpeg format
  --title TITLE         Window title
  --nodrop              If set, disable frame drop feature
  --repeat              If set, repeat video decoding
  --h265                If set, the specified video file will be assumed as
                        H.265. Otherwise, assumed as H.264
  --model MODEL         Model weight file
  --task TASK_DESC      Task description file
  --csv MAX_CSV_REC     Maximum CSV records
  --csvpath CSV_PATH    Directory path to save CSV files
  --verbose             If set, print debug message

For MIPI-CSI camera, use any negative number as the camera number.

$ python3 trt_pose_app.py --camera -1 

For USB Web camera, if you camera is detected as /dev/video1, use 1 as the camera number.

$ python3 trt_pose_app.py --camera 1

To get a CSV output, please use the --csv option with the maximum capture frames. You can specify the directory to hold the CSV file with the --csvpath option.

$ python3 trt_pose_app.py --camera 0 --csv 1000 --csvpath ./logs

To use the densenet121_baseline_att_256x256_B_epoch_160.pth pre-trained model which is also released at the resnet18_baseline_att_224x224_A model file, use the --model option.

$ python3 trt_pose_app.py --camera 0 --model densenet121_baseline_att_256x256_B_epoch_160.pth

This application can accept not only camera capture but also H.264 movie file as input. For high resolution movie file input, the --nodrop option might be needed.

$ python3 trt_pose_app.py --nodrop test.mov