Domain adaptation methods were briefly mentioned in Chapter 4, Influential Classification Tools, among transfer learning strategies. Their goal is to transpose the knowledge acquired by models from one source domain (that is, one data distribution) to another target domain. Resulting models should be able to properly recognize samples from the new distribution, even if they were not directly trained on it. This fits scenarios when training samples from the target domain are unavailable, but other related datasets are considered as training substitutes.
Suppose we want to train a model to classify household tools in real scenes, but we only have access to uncluttered product pictures provided by the manufacturers. Without domain adaptation, models trained on these advertising pictures will not perform properly on target images with actual clutter, poor lighting, and other discrepancies.
Training recognition models on synthetic...
