Driving in the Matrix: Can Virtual Worlds Replace Human-Generated\n Annotations for Real World Tasks?

摘要

Deep learning has rapidly transformed the state of the art algorithms used to\naddress a variety of problems in computer vision and robotics. These\nbreakthroughs have relied upon massive amounts of human annotated training\ndata. This time consuming process has begun impeding the progress of these deep\nlearning efforts. This paper describes a method to incorporate photo-realistic\ncomputer images from a simulation engine to rapidly generate annotated data\nthat can be used for the training of machine learning algorithms. We\ndemonstrate that a state of the art architecture, which is trained only using\nthese synthetic annotations, performs better than the identical architecture\ntrained on human annotated real-world data, when tested on the KITTI data set\nfor vehicle detection. By training machine learning algorithms on a rich\nvirtual world, real objects in real scenes can be learned and classified using\nsynthetic data. This approach offers the possibility of accelerating deep\nlearning's application to sensor-based classification problems like those that\nappear in self-driving cars. The source code and data to train and validate the\nnetworks described in this paper are made available for researchers.\n