How to pass an array to downstream element with python?

[nam1012.2016]

Please provide complete information as applicable to your setup.

• Hardware Platform : dGPU
• DeepStream Version: 7.0
• TensorRT Version: 8.6.1.6
**• NVIDIA GPU Driver Version: 555.58.02 **
• Issue Type: Question

For example, each object from pgie have additional data (for example an array like[1,2,3,4]) and this array is needed for preprocessing in sgie. How can i send this data to downstream elements?

[fanzh]

1. could you share the scenario? how are the additional data used n sgie?
2. what is the whole media pipeline? is it “…->pgie->nvdspreprocess->sgie->…”? you can add usermeta to objectmeta by nvds_add_user_meta_to_obj. please refer to this sample. then in nvdspreprocess, fetch and process these usermeta. please refer to this sample.

[nam1012.2016]

So i have a pipeline that look like this

pgie(face detect with landmark)-> nvdspreprocess(use the landmark to align face) -> sgie(face embedding)

Currently i got face landmark after pgie post processing but i dont know how to send it to nvdspreprocess.

[fanzh]

1. Thanks for the sharing! please refer to the two samples in my last comment. you can use nvds_add_user_meta_to_obj to add custom user data for every object. then in nvdspreprocess, you get the custom user data from the obj_user_meta_list.
2. will you generate new face or new landmark in nvdspreprocess? what the results of “use the landmark to align face”?

[nam1012.2016]

I actually solved this by allocate an array with mask_params as i have no intention of doing any instance segmentation task later.

# Add mask data at pgie src pad
    def modify_mask_data(obj_meta, landmark_data):
        # Access the mask_params from obj_meta
        maskparams = obj_meta.mask_params
        maskparams.height = 1
        maskparams.width = len(landmark_data)
        # Create a NumPy array to hold the mask data (e.g., a binary mask)
        mask_array = landmark_data  

        # Assuming alloc_mask_array allocates memory for the mask array
        mask_array_ptr = maskparams.alloc_mask_array()

        if isinstance(mask_array_ptr, np.ndarray):

            mask_array_ptr[:] = mask_array  # Copy the mask data to the allocated array
            #print("Updated mask array:", maskparams.get_mask_array())
        else:
            print("Error")

# Acess data in nvdspreprocess: 
  float* mask_data = obj_meta->mask_params.data;
  unsigned int mask_size = obj_meta->mask_params.size / sizeof(float); 
  for (unsigned int i = 0; i < mask_size; ++i) {
      float value = mask_data[i];
      printf("Landmark value: %f\n",value);
      // Process the mask value as needed
  }

Maybe i will have to create custom user data but this works for now. Thanks for your support, absolutely love it.