detect

“detect” filter is designed to load the models for object detection only. It checks model output layer according to the model_proc configuration file, converts the detection results to bounding boxes list, and keeps them in AVFrameSideData instances in the input frame’s AVFrame instance. Besides the coordinates of the bounding box, label id and the confidence of the detection result are also kept.

Filter detect
  Image Inference Detect Filter.
    Inputs:
       #0: default (video)
    Outputs:
       #0: default (video)
inference_detect AVOptions:
  dnn_backend       <flags>      ..FV..... DNN backend for model execution (default 1)
  model             <string>     ..FV..... path to model file for network
  model_proc        <string>     ..FV..... model preproc and postproc
  object_class      <string>     ..FV..... objective class
  device            <string>     ..FV..... running on device name
  configs           <string>     ..FV..... configurations to backend
  interval          <int>        ..FV..... detect every Nth frame (from 1 to 1024) (default 1)
  nireq             <int>        ..FV..... inference request number (from 1 to 128) (default 1)
  batch_size        <int>        ..FV..... batch size per infer (from 1 to 1000) (default 1)
  threshold         <float>      ..FV..... threshod to filter output data (from 0 to 1) (default 0.5)
  square_bbox       <boolean>    ..FV..... optimizing bbox for face detect (default false)
  crop_params       <string>     ..FV..... cropping rectangle format x|y|w|h
  async_preproc     <boolean>    ..FV..... do asynchronous preproc in inference backend (default false)