STAPLE: Siamese Transformer Assisted Pseudo Label Ensembling for Unsupervised Domain Adaptation in No-Reference IQA

Arshita Gupta, Zhe Zhu, Tien Bau; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 769-778

Traditional Image Quality Assessment (IQA) models rely on image - Mean Opinion Scores (MOS) pairs from a specific source domain to train a robust mapping function in a fully supervised manner. While these methods excel on test images with distortions similar to the training data, their effectiveness drops sharply when confronted with novel distortions, necessitating expensive fine-tuning with additional annotations and hyper-parameter optimization. To address these challenges, we present Siamese Transformer Assisted Pseudo Label Ensembling (STAPLE), a novel Unsupervised Domain Adaptation (UDA) technique for No-Reference IQA. STAPLE leverages a Siamese Transformer Network (STN) that learns quality differences between image pairs and generates high-quality pseudo-labels by pairing unlabeled target images with labeled source images. By ensembling predictions from multiple source references, our method robustly reduces variance and consistently maintains high accuracy on both source and target domains. Notably, STAPLE excels in both synthetic to in-the-wild and in-the-wild to synthetic scenarios, even when there is zero overlap between the distributions. Extensive experiments on multiple benchmarks confirm that STAPLE enhances performance in challenging, real-world conditions.