Augmented Reality Meets Computer Vision : Efficient Data Generation for Urban Driving Scenes

How can we create large training sets for deep neural networks while avoiding labor intensive annotation of real images or expensive creation of synthetic content? We suggest near photo-realistic augmentation of real images with synthetic objects to combine the best of both worlds. The extended version of our BMVC’17 paper investigates which aspects matter most in the context of object detection and instance segmentation in urban driving scenes: Arxiv link

Posted in arXiv.org.