Fig. 3
From: Pest detection in dynamic environments: an adaptive continual test-time domain adaptation strategy
The proposed dynamic data augmentation. The image \({{\textbf {X}}}\) undergoes augmentation by a strong augmenter and is classified by a classifier to compute the binary cross-entropy loss \(\mathcal {L}_{bce}\), which updates the classifier. Batch difference loss \(\mathcal {L}_{bdf}\) and domain difference loss \(\mathcal {L}_{dif}\) are calculated and combined to update the strong augmenter. After training, the last 3 rounds of images form the strong-augmented dataset. The input image is also augmented by a weak augmenter for 3 rounds, tripling the dataset size to create the weak-augmented dataset. Both augmented datasets and the original images are used as inputs for the model