Realization:
A CNN network was created that was resistant to the detection of various types of masks: respirators, gauze, medical. As well as their color: motonone, black, pink and other colors, as well as patterned. The CNN network has been optimized, the dimension has been reduced to INT8, and pruning has also been performed. The network is ported to Xilinx via VitisAI.

140% fps
>85% Reduced the size of neural network
Value:

• A full-fledged application for Zynq UltraScale+ MPSoC chips has been created
• The solution is implemented in the customer's smart cameras

Project:
Creating a binary classifier based on the ML algorithm. Porting it to the KU060 chip by means of VitisAI. Integration with a C++ application camera controls on the satellite.

700Tb Satellite images analyzed

What we have done:

• We assembled and marked up a turnkey dataset.
• We trained the yolo network on our own GPU resources.
• Conducted pruning.
• Integrated the network with the automotive suspension control application of a large automaker.

Project:

• Development of a lightweight version of a neural network for detecting objects (people, vehicles, animals).
• Portation of the neural network of object detection on the FPGA Lattice.
• Creation of an algorithm for calculating a dangerous approach based on
• data about objects received from a neural network.

Value:
The client received a turnkey software and hardware solution for mass production based on an affordable AI chip of the Lattice family.

Control via video surveillance
The fight detector allows you to automatically detect non-standard behavior of people

Identification of socially dangerous actions
Hooliganism, quarrels, mugging and scuffle

Without human involvement
The service works with multiple cameras at the same time and does not require an operator in the monitoring center

Project:
To count calories in the customer's fitness application, it was necessary to make an autonomous recognition of food, fruits, vegetables and drinks, which can be integrated as an SDK.

Realization:

• A hierarchical CNN network has been created
• Optimized to work on iPhone
• Recognizes more than 1000 food dishes
• Works as an SDK via library import.

Value:
The client will receive a unique detuning from competitors and has implemented new functionality in his application.