EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems
Researchers have introduced EmbodiTTA, a novel framework for resource-efficient test-time adaptation (TTA) designed for embodied visual systems on edge devices. This approach, termed on-demand TTA, activates adaptation only when a significant domain shift is detected, thereby reducing computational overhead. EmbodiTTA incorporates a lightweight domain shift detection mechanism, a source domain selection module for robust accuracy, and a decoupled Batch Normalization update scheme for memory-efficient adaptation with small batch sizes. AI
IMPACT Enables more efficient and practical deployment of adaptive AI models on resource-constrained edge devices.