Self-supervised Learning of Object Landmarks from Unlabelled Data

Tomas Jakab (University of Oxford, UK)

Abstract:

Modern machine learning methods can solve complex image labelling tasks with good accuracy but at the cost of collecting large annotated datasets for training. The cost of these manual annotations is a major obstacle to deploying machine learning to new tasks. Removing the annotation bottleneck is thus one of the key objectives of current research in computer vision.

In this talk, we will explore the problem of learning the 2D geometry of object categories from unlabelled images or videos. Specifically, I will introduce our novel method that learns to discover object landmarks without any manual annotations. It automatically learns from images or videos and works across different datasets of faces, humans, and animals. I will also show how a weak empirical prior can be incorporated into the method in order to learn human-interpretable landmarks.