One Model to Translate Them All: Universal Any-to-Any Translation for Heterogeneous Collaborative Perception
Researchers have developed UniTrans, a novel universal model designed for any-to-any feature modality translation in collaborative perception systems. This model addresses the challenge of heterogeneous sensor data by pre-training a set of translator experts and dynamically combining them for new modality mappings. UniTrans achieves zero-shot translation by extracting scene-invariant codes from intermediate features, outperforming existing methods on benchmark datasets and offering a scalable solution for real-world applications. AI
IMPACT Enables more efficient and scalable fusion of diverse sensor data in collaborative perception systems.