Zero-label highway driving
Status: Deprecated in favor of unsupervised learning from video frame prediction (v1 project). Other teams trying supervision from physics found it challenging as well. But learned a lot.
Target abstract: We demonstrate that it is possible to train an object detector for autonomous driving applications using zero labels, only the laws of physics. The detector achieves [very high]% MAP when trained on the Kitti self-driving dataset, but is trained entirely without labels.
See https://docs.google.com/document/d/1sd4o7wGOfrtj8S5-IOkSbRkI-my4lBqVHVMrVuNBv50/edit for a high-level paper outline and priority list. See https://docs.google.com/document/d/1DdLuUGetVyXf1WBOIEo2owMpZbl12arFeOvmcmBQnWg/edit for the hour-by-hour experiments log.