The ubiquity of mobile devices makes them a practical option for machine learning applications like recognition and detection. However, they suffer from memory and compute limitations, which are aggravated by resource-hungry deep learning models. Machine Learning for mobile devices has been an active research topic. In the case of mobile object detection, Single Shot Detector(SSD) and You Look Only Once (Yolov5) are the popular detectors, which are optimized for high fps and low memory footprint. Nevertheless, training such models, still require a good amount of labeled data. Depending upon the use case, it might be difficult to obtain and costly to annotate such data. In order to overcome this aspect, I will present how synthetic images can be useful with Domain Adaptation and Pseudo Labeling techniques.

Download the slides for this talk.Download ( PDF, 9576.54 MB)