Modern deep convolution neural networks have been responsible for a large increase in performance for tasks such as image classification. Recent work has shown that it is also possible to use these models to reason about geometry in images. This in turn allows us to estimate, for each pixel, the depth of each object in a given scene. Unfortunately, these existing methods rely on having large quantities of reliable ground truth data to train on, which can be challenging to obtain in large quantities.
In this talk I will present our recent work on pixel level depth estimation in images. By posing depth estimation as an image reconstruction problem we are able to learn how to generate depth without requiring any ground truth data. I will conclude by showing results on vehicle mounted video cameras illustrating that low cost cameras have the potential to be used as cheap depth sensors. A video of our results can be seen here https://www.youtube.com/watch?v=v8cpDQ22bSg
Project Website and Video: