Self-supervised fine-tuning of dense computer vision models for autonomous driving
||Self-supervised fine-tuning of dense computer vision models for autonomous driving|
||Adam Tonderski <email@example.com>|
||2022-02-03 – 2022-09-01|
Self-supervised learning is a rapidly developing field of research that has shown great success in natural language processing. In the domain of computer vision progress has been slower, but recently self-supervised pre-training has started to outperform supervised pre-training on large datasets. However, much of this gain comes from using immense amounts of both data and computational resources.
In this project we study how such general-purpose self-supervised models can be adapted to a specific domain - here autonomous driving. Specifically we start off with a checkpointed model from for example imagenet, and fine-tune it in a self-supervised manner on autonomous data. This model is then evaluated in several downstream supervised tasks, such as object detection. The aim is to improve this intermediate self-supervised step so that the fine-tuning performance on the down-stream tasks is maximized.