| Summary: | Object parts, also known as landmarks, convey information about an object’s shape and spatial configuration in 3D space, especially for deformable objects. The goal of landmark detection is to have a model that, for a particular object instance, can estimate the locations of its parts. Research in this field is mainly driven by supervised approaches, where a sufficient amount of human-annotated data is available. As annotating landmarks for all objects is impractical, this thesis focuses on learning landmark detectors without supervision. Despite good performance on limited scenarios (objects showcasing minor rigid deformation), unsupervised landmark discovery mostly remains an open problem. Existing work fails to capture semantic landmarks, i.e. points similar to the ones assigned by human annotators and may not generalise well to highly articulated objects like the human body, complicated backgrounds or large viewpoint variations.
In this thesis, we propose a novel self-training framework for the discovery of unsupervised landmarks. Contrary to existing methods that build on auxiliary tasks such as image generation or equivariance, we depart from generic keypoints and train a landmark detector and descriptor to improve itself, tuning the keypoints into distinctive landmarks. We propose an iterative algorithm that alternates between producing new pseudo-labels through feature clustering and learning distinctive features for each pseudo-class through contrastive learning. Our detector can discover highly semantic landmarks, that are more flexible in terms of capturing large viewpoint changes and out-of-plane rotations (3D rotations). New state-of-the-art performance is achieved in multiple challenging datasets.
|
|---|