FastestStreamDet最速流推理

A inference implementation that supports multiple input streams in a single file based on FastestDet.

一个基于FastestDet的单文件多类型输入推理实现

开箱即用兄dei

Advance强处

• Use the ONNX model to infer real-time camera streams, images, and video file streams.
• 使用通用开源的onnx模型格式，能推理摄像头实时流、图片、视频文件流等多种类型的输入。
• Simply output inference results in a variety of ways.
• 输出可随心所欲地以不同方式（保存、字节码导出返回……）处理，文件里都给你提前写好想要什么输出方式了
• Single file.
• 单文件。（指的是命令行用的核心部分，UI实现看看行不行吧）
• Debugged problem carefully, fixing most bugs.
• 踩了无数的坑,包没问题的,有错你踹Issue我

How to use咋用

Tips: Never use ctrl+c to force an working program stopping. You may found the camera still being used.

别手贱直接CTRL+C给正常工作的推理进程干掉了，不然用摄像头的时候有你好果子吃的（摄像头：这给我干哪了，这还是地球吗？）

Install the required dependencies firstly

先把下面这些库装好,python不是3.8版本的话会有奇妙的bug

The dependency onnxruntime might be hard to installed on Linux platforms.

Linux装onnxruntime是真的坐牢，慎用

• Python(3.8 for recommended version) and pip
• opencv-python
• numpy==1.23.0
• onnx_simplifier==0.3.10
• onnxruntime==1.16.0
• pathlib
• argparse

Demo示例

• Type in the terminal:

• 复制到终端回车：

  python detect.py --source datatest/6.jpg
  

• Check your result image in the resultsave folder.

• 看看 resultsave 文件夹里保存的推理结果

• Type in the terminal:

• 复制到终端回车：

  python detect.py --source video --videofile datatest/21115-315137069_small.mp4
  python detect.py --source video --videofile datatest/2174-155747455_small.mp4
  

• The First example video owned by Jahhoo

• The Second example video owned by RafterJr72

• 以上示例视频由这两位作者分别所拥有

• Check the detect result through the new window created by opencv

• 通过opencv打开的视频流窗口查看推理结果

• Click "q" to end the detect process.

• 按“q”键终止推理进程

CameraStream摄像头实时流

• Choose your own camera index then type in terminal:

• 选好摄像头的编号复制到终端回车：

  python detect.py
  

• or use your own camera index:

• 不然就在下面的命令后边打上摄像头在系统中的序号，Windows是设备管理器，Linux可以用v4l2-ctl：

  python detect.py --source {your camera index, it usual in integer type}
  

• Check the detect result through the new window created by opencv

• 通过opencv打开的视频流窗口查看推理结果

• Click "q" to end the detect process.

• 按“q”键终止推理进程

Picture图片

• Type in the terminal:

• 复制到终端回车：

  python detect.py --source {your Picturefilepath}
  

• Check your result image in the resultsave folder.

• 看看 resultsave 文件夹里保存的推理结果

VideoStream视频文件流

• Type in the terminal:

• 复制到终端回车：

  python detect.py --source video --videofile {your videofilepath}
  

• Check the detect result through the new window created by opencv

• 通过opencv打开的视频流窗口查看推理结果

• Click "q" to end the detect process.

• 按“q”键终止推理进程

Todo待完成

• Save the VideoStream detection result as a file
• 以文件形式保存视频流推理结果
• Write an UI to adapt most usage
• 写一个能适用于大多数场景需求的UI
• Multi CPU core Inferences
• 多核推理
• Use ncnn Framework(almost done)
• 使用ncnn来编译推理(接近完成,孩子原先的ncnn已经编译成功啦)

Reference引用

• FastestDet
• onnx
• Jahhoo
• RafterJr72