How to use the object detection model trained by eiq on imx8mp board or pc?

[taotaowang]

Hello,

I trained a model according with "eIQ_Toolkit_UG.pdf" followed below routine. 

import my dataset -> select model -> detection -> balanced -> npu -> trainer -> validate -> deploy -> export model -> tensorflow lite model saved.

Below is the validate result. 

below is the export model sumarry.

Model property is like below.

As upper properties display, this model output name is StatefulPartitionedCall and type is float32[1,2034,7] , But where can i get the output_data defination?  How can i use this model in my inference program?

In my inference program i used below codes to get the inference result but it seems not correct.

  output_data = interpreter.get_tensor(output_details[0]['index'])

  print(output_data)

  for i in range(output_data.shape[1]):

        score = output_data[0][i][0]

        class_id = output_data[0][i][1]

        x_min = output_data[0][i][2]

        y_min = output_data[0][i][3]

        x_max = output_data[0][i][4]

        y_max = output_data[0][i][5]

        other = output_data[0][i][6]

Please give me some advises about the eiq detection model inference realization.

Thanks !

Taotao Wang

[jake4]

Hi @Zhiming_Liu 

In the  eIQ Documentationcouldnt find this information.

We use the eiqtool for training a object_detection model with mobilenet ssd v3 for a imx8mp NPU.

The example given in eIQ_Toolkit_v1.16.0\workspace\models\mobilenet_ssd_v3\mobilenet_ssd_v3.ipynb

the inference is done by using the RTview and tensorflow, which our imx8mp we dont want to include . We would like to use the tensorflow lite inference.

We get a output of tensor (Since I have 1 class ) 

name: StatefulPartitionedCall:0

tensor: float32[1,2034,6]

location: 392

By splitting into scores and bounding box for each (1, 2034, 2), (1, 2034, 4)

And then follow the output treatment as in mobilenet_ssd_v3.ipynb, we dont get the bounding boxes and the scores as expected.

So wondering, whats the output format signature is?

We could guess (from https://community.nxp.com/t5/eIQ-Machine-Learning-Software/How-to-interpret-the-output-from-a-mobile...) 

The model predicts 2034 detections per class. The [1,2034,4] tensor corresponds to the box locations in terms of pixels [top, left, bottom, right] of the objects detected.

And [1,2034,2] tensor corresponds to scores of our class and background.

We couldnt get a correct meaningful output from the model trained by the eiq tool, which seem to follow a unique way of combining the outputs bounding boxes, that can be only interpreted by the rtview engine. But can we have a detailed explanation of the output. And example to show without using rtview or tensorflow libs?

Thanks.

[Zhiming_Liu]

Please refer the eIQ document with HELP-->eIQ Documentation