DeepView IoT Sensor
Frequently Asked Questions
What if my question is not in the list below
Contact us and we will answer your question and add it to this list
How is the IoT vision sensor micro different from an industrial smart camera?
The IoT Vision Sesnor micro is intended for OEM applications that require an optimized integrated solution suitable for volume production. The core hardware elements of this design have a Bill of Materials (BOM) cost of less than $50 in moderate volume enabling many applications not practical with higher cost modules or single board computers. To achieve this lower BOM cost point the sensor resolution and processing speed is much lower than many smart camera designs based on high performance application processors, GPU’s or NPU’s. For many IoT sensor applications this trade-off is reasonable as the sensor must also be suitable for lower power operation.
What is the difference between a traditional vision sensor and deep learning-based vision sensor?
A traditional vision sensor is programmed with rules-based algorithms to perform a specific function such as bar code reading, light sensing, or distance measurement to a surface. Using deep learning a smart vision sensor can be trained to detect or classify specific objects or their characteristics enabling more complex use cases.
What is the difference between Cloud and Edge Inference and why does it matter?
First generation IoT sensors typically have sent raw results to the cloud for further processing and analysis. This is appropriate for simple sensors or vision sensors with continuous high bandwidth cloud connectivity. For many vision sensor applications on device edge processing offers some clear advantages in deterministic processing latency, lower communication costs, and greater power efficiency. With the DeepView IoT Vision Sensor system architecture the image processing and deep learning inference is performed on the device and only the result reported to the local application or cloud.
How does a Smart Camera module compare to an Integrated Vision Sensor?
Some Smart camera modules contain on board processing which may be configured for running edge deep learning models with high performance and good power efficiency. Many of these designs require an additional single board computer to implement the balance of the system for remote operation which should be considered when selecting an architecture.
How many images are required to train the IoT Vision sensor for my detection or classification use case?
This varies with the specific deep learning model and application requirements. Typically, a standard pre trained classification model can be customized with hundreds of images to start to provide good results if the images well represent the range of real-world conditions. Adding thousands of training images over time will improve the accuracy and robustness of the solution
Can the Deep Learning Model be updated in the field?
Yes, in additional to full firmware updates the DeepViewRT engine supports full or partial model updates which have been optimized for runtime size to minimize the memory requirements and communication bandwidth which can be significant for high volume deployments.