Unsupervised domain adaptation (UDA) has become increasingly prevalent in scene text recognition (STR), especially where training and testing data reside in different domains. The efficacy of existing UDA approaches tends to degrade when there is a large gap between the source and target domains. To deal with this, gradually shifting from domain to domain is the key. We introduce the Stratified Domain Adaptation (StrDA) approach, which examines the gradual escalation of the domain gap during learning: it partitions the target data into subsets so a progressively self-trained model can adapt to gradual changes, using domain discriminators to estimate the out-of-distribution and domain-discriminative levels of each sample. Extensive experiments on benchmark scene-text datasets show that our approach significantly improves baseline source-trained STR models.
Stratified Domain Adaptation (StrDA) is an unsupervised domain-adaptation method for scene text recognition (STR). When the gap between the source (training) and target (test) domains is large, existing UDA methods degrade; StrDA instead shifts gradually, domain by domain.
The key idea is to partition the unlabeled target data into ordered subsets of increasing domain gap, then progressively self-train the model so it adapts to gradual change. StrDA measures each sample's proximity to the source and target domains using domain discriminators that estimate out-of-distribution and domain-discriminative levels, and uses these estimates to stratify the data for staged self-training.
On standard scene-text benchmarks — six core benchmarks plus additional datasets — StrDA significantly improves the performance of baseline (source-trained) STR models across multiple recognizer architectures (e.g. CRNN, TRBA, ABINet). The method appears at WACV 2025 (IEEE/CVF Winter Conference on Applications of Computer Vision); source code is publicly available.