Make a parsing function for person attribute recognition classifier

I am using the model called duke in this repo it’s a person attribute recognition model, I’ve convert the model to tensorRT and now try to parse the output of that model, Could you recommend a way to parse that model output, specially adding the labels at the parsing function ??
and could i get the tensorMeta output from a classifier ?

• (GPU)
• DeepStream Version: 5.0
• Driver Version: 460.80 )
• Issue Type( new requirements)

Yes. You can get tensor meta output from any model with config gst-nvinfer “output-tensor-meta=1” and “network-type=100”.

Please refer to:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html#gst-nvinfer-file-configuration-specifications

https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html#tensor-metadata

1 Like

This solved my question, thank you ^^
but I need now to make a pre-processing for each object before the secondary model classify them, pre-processing such as:

  • Resize each object to (288, 188)
  • Normalize the objects with this formula T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), but i dnt know how to make this using the sgie ??

Unfortunately, current gst-nvinfer dose not support different deviation for different channel. You can only set one deviation for all channels by “net-scale-factor”.

The mean values can be set by “offsets”.

Please read the document carefully. Gst-nvinfer — DeepStream 6.1.1 Release documentation

All parameters are explained in details.

1 Like

Thank you @Fiona.Chen
but what about resizing an object before fed it to the classifier ?

The gst-nvinfer already knows the model dimensions by “infer-dims”, so it will scale the input to the model dimensions.

@Fiona.Chen could you check this new topic, it’s related to this one

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.