Efficient Test-Time Adaptation for Dynamic Distribution Shifts

Pre-trained visual models are increasingly deployed in streaming environments where the target distribution changes with weather, illumination, sensor aging, background composition, and time. Existing test-time adaptation (TTA) algorithms usually assume either a fixed target domain or sufficiently large independent test batches; consequently, they can be computationally heavy, slow to converge, and vulnerable to noisy pseudo-supervision under dynamic shifts. This paper presents an efficient online TTA framework for dynamic distribution shifts. The framework contains dynamic confidence screening, branch-level incremental tuning, and temporal distribution smoothness. Only a small set of adaptive branches is updated during inference, while unreliable samples are filtered before optimization and historical batch statistics regularize short-term drift. Experiments on dynamic image classification and object detection streams show average accuracy gains of 4.2%–8.7% over representative TTA baselines and more than 60% lower inference latency than full-parameter continual adaptation. The proposed design provides a practical balance between robustness, stability, and real-time deployment efficiency. Unlike prior TTA designs that treat filtering, lightweight updating, or temporal stabilization separately, the proposed framework couples reliability-dependent sample admission, residual branch-only updating, and statistics-level temporal memory within the same fixed online update budget. Because the classification corruption streams are manually scheduled, the evidence is interpreted as controlled dynamic-benchmark validation rather than a substitute for field deployment.