CVPR Daily - Tuesday

Riccardo says he was inspired by similar works, including the MAML algorithm for meta-learning applications , which the team have built on for their meta- learning-based solution. “ The first work that proposed domain randomization to improve the generalization of computer vision models was also a huge source of inspiration, ” he adds. “ It essentially defined domain randomization as randomizing the rendering of the samples from a simulator to learn models for robotics that will be less surprised in the real world. By randomizing many aspects of your training samples, the model will be better prepared because by randomizing, you increase your training dataset more and more. We work out of simulation, with input image datasets, but we do domain randomization by applying random image transformations to randomize our training set more and more . This helps a lot in this continual domain adaptation problem. ” He says a neat feature he would like to include is one that would make smarter decisions on how to run the model training set . Domain randomization is achieved by applying random image transformations, but there are no specific rules on how to choose these image transformations because you don ’ t know on which new domains you will deploy the model. Riccardo ’ s co-authors are Diane Larlus and Grégory Rogez , who both work with him at NAVER LABS Europe . Diane is a principal investigator and Grégory leads the computer vision team. “ A big thank you to my great co-authors, ” he beams. “ It was a great pleasure working with them. ” 11 DAILY CVPR Tuesday Riccardo Volpi