Self-supervised multi-task learning for self-driving cars

©istock

This project aims to improve automatic learning techniques for autonomous vehicle systems.

For an autonomous vehicle to operate, it must first be taught to drive. The technique used is machine learning, which consists in teaching the “brain” of the car to recognize everything around it. To do this, it is provided with a large number of images of multiple possible situations. The more there are, the better. But to be effective, learning requires that these images be annotated manually to explain to the car’s brain what they represent. A long and tedious job.

The objective of this project is therefore to reduce the need for human annotations by developing a self-supervised learning model. In particular, the use of computer-generated images makes it possible to obtain the corresponding annotations automatically. However, the learning process must then be modified before real images can be used instead of the generated ones. In this context, the Visual Intelligence Laboratory for Transportation (VITA) will work on new training strategies for unsupervised domain adaptation that use additional privileged information on the synthetic domain during training to improve transfer to the real one.

Principal investigator Prof. Alexandre Alahi
Sponsor Samsung
Period 2019-2021
Laboratory VITA