BiFine: Bilateral Fine-Grained Alignment with Dual Channels for Partial Domain Adaptation

EasyChair Preprint 13383

Abstract

Partial Domain Adaptation (PDA) often grapples with negative transfer when the target label space is a subset of the source domain's. Addressing this, we present BiFine, a dual-channel adversarial weighting framework for PDA that orchestrates a bilateral fine-grained alignment between domains. The Dual-Channel consists of two key components: the Shared-Private Weighting Diverger (SPW) and the Centroid-Based Similarity Discriminator (CSD). The SPW selectively modulates weights for shared classes, amplifying them to enhance positive transfer while suppressing those potentially leading to negative transfer from private source domain classes. Concurrently, CSD employs a bilateral strategy by adjusting target sample weights based on their cosine similarity to the centroids of shared source classes and attenuates intra-class variances to sharpen class boundaries. This holistic approach promotes a refined domain adaptation, securing closer alignment for shared classes and segregating outliers. Extensive evaluations on ImageCLEF, Office-31 and Caltech-office datasets affirm BiFine's efficacy, outperforming exsiting methods with classification accuracies of 91.99%, 97.78% and 96.49%, respectively.

Keyphrases: Dual-Channel Weighting, Fine-grained Alignment, Partial domain adaptation, negative transfer

@booklet{EasyChair:13383,
  author    = {Zhongze Wu and Yitian Long and Shan You and Xiu Su and Jun Long and Yueyi Luo and Chang Xu},
  title     = {BiFine: Bilateral Fine-Grained Alignment with Dual Channels for Partial Domain Adaptation},
  howpublished = {EasyChair Preprint 13383},
  year      = {EasyChair, 2024}}