R&D company DANNIE LT announced a successful implementation of a segmentation algorithm on Ingenic T40. T40 is a specialized for computer vision SOC with hardware NN acceleration.

Segmentation algorithm determinates exact object boundaries on an image frame, compared to boundary box algorithm.

This allows to solve more complex business tasks like a type of road’s surface detection on cost effective and energy efficient (<1W) T40 platform, that has been never done before.

“Working with the top level engineering team from Ingenic was a great experience and honor for us. I am looking forward to new achievements we can do together“, ― Viktor Vikulov, head of CV department in Dannie LT.

The whole project took about 3 month and was not possible without tight collaboration with Ingenic R&D team. Memory limitations, hardware acceleration specifics and completely new functionality for NPU sdk – great challenges we had to overcome together.

“It was high risk-high reward R&D project. But now, we can provide our customers with highly competitive solutions for cost sensitive and energy efficient AI products», ― Alex head Dannie Lt R&D department.

For model Dannie Ml engineers used customized resnet and u-net with Int8 quantization. There are still some work needs to be done in field of performance, but even current result allows us to implement ADAS features on personal electric vehicles for our customer.