EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems
EmbodiTTA introduces on-demand Test-Time Adaptation (TTA) for resource-constrained devices by adapting models only when a significant domain shift is detected. It uses three techniques: lightweight domain shift detection, source domain selection, and decoupled Batch Normalization updates. This framework achieves high accuracy while drastically reducing energy consumption and computation overhead compared to existing TTA methods.