What, How, and When Should Object Detectors Update in Continually Changing Test Domains?

We provide additional implementation details for each baseline model. Our framework incorporates all baseline models using the official code except Mean-Teacher. The results of the experiments are reported based on the optimal hyperparameters that yield the best results in our scenario. ActMAD [31] As ActMAD exclusively conducts experiments on the KITTI dataset, where all images have a constant height and width (e.g., 370 x 1224), ensuring consistent feature map sizes for all samples. ActMAD can easily align them along the spatial axis. However, in the general setting of object detection tasks, such as the COCO benchmark set, where image sizes and width-to-height ratios vary, aligning feature maps along the spatial axis becomes challenging due to different sizes. To adapt ActMAD to our COCO ! COCO-C scenario, we perform center cropping on the feature maps to match the size of training domain feature maps and the current test sample feature maps. We employ a learning rate of 1e-5 for COCO and 1e-4 for SHIFT, respectively. Mean-Teacher As the official code of TeST [38] is not available, we implement the EMA-updated Teacher and Student models following TeST [38], to conduct experiments in our scenarios. TeST involves three forward steps for a batch: forwarding weakly augmented samples through the student network, strong augmented samples through the teacher network, and original samples through the teacher network for outputs. However, for a fair comparison, we perform two forward steps, forwarding the original sample through the teacher network and strong augmented samples through the student network, to make predictions before adaptation for every samples. We utilize a learning rate of 1e-5 and set the EMA update rate for the teacher network to 0.999. NORM [37] We set the hyperparameter N that controls the trade-off between training statistics and estimated target statistics as 128. DUA [30] We set the momentum decay as 0.94, minimum momentum constant as 1e-4, and the initial momentum decay as 1e-3.