Research Focus
My current research focuses on backward-compatible representations—learned features that can be used interchangeably between models without wasting training data as architectures evolve. This is a natural evolution of my long-standing interest in incremental learning: from early research in multi-view geometry and 3D reconstruction, to incremental camera tracking (SLAM with zooming cameras) and object tracking in lifelong video streams. Object tracking, particularly appearance learning, later evolved into incremental appearance learning studied through continual learning with deep neural networks. This phase brought the challenge of catastrophic forgetting, and the advent of foundation models reshaped the challenge into finding ways to obtain representations that remain compatible across models. Throughout these stages, my objective has remained consistent: to develop vision systems and AI methods capable of incremental lifelong learning, ensuring that learned representations remain compatible while retaining prior knowledge.