IMX500 intelligent vision sensorIMX500 intelligent vision sensor

Transform your business with IMX500 Intelligent Vision Sensor

Develop vision based AI applications using a high speed edge processing one-chip solution

Convert

Convert your own quantized and trained deep neural network model into an optimized binary file, ready to package and deploy on IMX500.

We support:

  • TensorFlow Lite with FlatBuffer
  • TensorFlow with frozen graph serialization, and a fixed input shape in the PlaceHolder

IMX500 developer tools include a converter to run on your own infrastructure or in the Smart Camera Managed App on the Azure cloud.

The converter includes a simulator which evaluates the neural network’s performance on the SDSP and generates a KPI report.

The converter can process a DNN and disregard the specific coefficient values. In this case, the network can be quantized, in floating point, or even not trained at all. If the required conversion is intended for deployment, however, then the input network must be trained and quantized to 8 bits uniform quantization.

You can deploy any type* of neural network model or, to get up and running even sooner for faster prototyping and evaluation, you can start with one of our pre-compiled networks:

  • Object detection: MobileNet SSD v1 (COCO)
  • Image classification: MobileNet v1

* Only inference graphs are supported in this version of the converter. Recurrent graphs are not supported yet.