Tuesday, 19th March 2024 Enhancing 2D Representation Learning with a 3D Prior - Mehmet Aygun Abstract: Learning robust representations of visual data is crucial in computer vision. Traditionally, this requires expensive labeled data. Self-supervised learning aims to overcome this by learning from raw visual data alone. However, most current methods focus on 2D images, lacking rich 3D information like humans. We propose enhancing self-supervised methods by integrating a strong 3D structural prior during training. Our experiments across various datasets show that our approach yields more robust representations compared to conventional self-supervised methods. Event type: Workshop Date: Tuesday, 19th March Time: 11:00 Location: G.03 Speaker(s): Mehmet Aygun Chair/Host: Bryan Li This article was published on 2024-11-22
