DeepStream SDK FAQ

Accelerated Computing Intelligent Video Analytics DeepStream SDK
37

29.[DSx_All_App] Debug Tips for DeepStream Nvinferserver Accuracy Issue
This corresponds to the Debug Tips for DeepStream Accuracy Issue and teaches you how to set relevant parameters in nvinferserver.

29.1 Input scale & offset

infer_config {
   preprocess {
     normalize {
        scale_factor: 0.0039215697906911373
        channel_offsets: [0, 0, 0]
     }
   }
}

29.2 Input Order

infer_config { 
  preprocess {
    tensor_order: TENSOR_ORDER_LINEAR
    network_format: IMAGE_FORMAT_RGB
  } 
}

29.3 Dims
Each model also needs a specific config.pbtxt file in its subdirectory for nvinferser. We set the imputlayer and outputlayer dims in the file.

input [
  {
    dims: [3, 224, 224]
  }
]
output [
  {
    dims: [6, 1, 1]
  }
]

For the nvinferserver config file:

infer_config { 
  backend {
    inputs: [ {
      dims: [3, 224, 224]
    }]
  }
}

29.4 scale and padding

infer_config { 
  preprocess {
    maintain_aspect_ratio: 1
    symmetric_padding: 1
  } 
}

29.5 inference precision
For the config.pbtxt file:

input [
  {
    data_type: TYPE_FP16
  }
]

For the nvinferserver config file:

infer_config { 
  backend {
    inputs: [ {
      data_type: TYPE_FP16
    }]
  }
}

29.6 threshold

postprocess {
  detection {
    nms {
      confidence_threshold:0.2
      topk:20
      iou_threshold:0.5
    }
  }
}