Self-supervised learning (SSL) and foundation models have advanced 3D perception for LiDAR and radar, but current models are trained independently, leading to fragmented representations.
- During your thesis you will develop a multi-teacher knowledge distillation framework to merge multiple pretrained models into a single, compact backbone, without labeled data.
- You will focus on task-agnostic distillation objectives to align heterogeneous feature spaces and prevent negative transfer, with extensions to integrate supervised teachers trained on unknown tasks. The resulting model will aim to deliver stronger, more general representations, reduced model capacity, and compatibility with existing perception stacks.
- Furthermore, you will conduct structured evaluations using both public datasets and internal real-world data, striving for excellent theoretical and practical results.