Cross domain adaption for CNN's-learning in a new environment

CNNs have been successful to a great extent in different computer vision tasks, but their efficiency is usually considerably lowered when they are transferred to new domains that vary in data distributions. This problem is called domain shift. In order to solve this, we come up with a convenient cross-domain adaptation framework that opens up the possibility for CNNs to be generalizable over dissimilar domains visually and that does not stipulate the involvement of labelled data. Our approach leverages a lightweight yet sufficiently robust backbone of EfficientNet-B0 and combines the Maximum Mean Discrepancy (MMD) technique for feature distribution alignment, and a confidence-based pseudo-labeling strategy for progressive classification in the target domain, and all of this together results in a very effective methodology. Relative to existing domain adaptation approaches, our model shows a remarkable benchmark performance on the Office-31 dataset (Amazon → Webcam) and on transfer tasks (DSLR → Webcam, Amazon → DSLR). The experimental evaluation and visual inspection of the results demonstrate that it is not only that our proposed method improves the accuracy of the target domain, but is also able to capture deep embedded representations that are invariant to domains. This framework can be applied to real-world settings due to prohibitively high costs or time spent to generate labeling.