Computer vision – object detection, localization, classification, and recognition. In this assessment period, following the global trend, our focus in computer vision has shifted to machine learning-based approaches, including deep convolutional NN, recurrent NN (e.g. LSTM), and generative adversarial networks. We tackle current obstacles in using supervised AI methods more broadly and research acquisition of annotated training datasets by faster labelling methods and generation of synthetic data (in collaboration with leading EU researchers from UNIMORE, TTTech, VIF, etc. in international projects such as H2020 PRYSTINE, Comp4Drones, etc.). Furthermore, we develop efficient computer vision algorithms for use on devices with limited computation resources;
In the embedded intelligence field we focus on computation and energy efficiency and develop novel architecture solutions for implementation of data processing algorithms (transformed ANN architectures etc.) on FPGAs and heterogeneous SoC devices (in collaboration with TECNALIA, Infineon, BUT, etc. in H2020 3Ccar, Autodrive, ENACT and other projects). This research results in enabling demanding perception processing (e.g. stereo-image processing and NN-based inference) on edge computing devices. Our expertise in embedded intelligence resulted in award from LZA for one of the most significant achievements in Latvian science in 2018 (work “An original approach for transforming the architecture of ANN into Field-programmable gate array”);