Deep networks are prone to performance degradation when there is a domain
shift between the source (training) data and target (test) data. Recent
test-time adaptation methods update batch normalization layers of pre-trained
source models deployed in new target environments with streami