Free Lunch to Meet the Gap: Intermediate Domain Reconstruction for Cross-Domain Few-Shot Learning

TL;DR

This work tackles Cross-Domain Few-Shot Learning by introducing Intermediate Domain Proxies (IDP) that reconstruct target-domain features from a source-derived codebook, enabling fast, data-efficient domain alignment through normalization-layer transformations. The approach combines dense feature reconstruction with a clustered intermediate proxy pool and a BN-statistics-based alignment mechanism, optimized via a triad of losses and trained in three stages (source pretraining, target finetuning, and target inference). The method demonstrates state-of-the-art performance across eight cross-domain benchmarks and provides theoretical and empirical evidence that intermediate proxies reduce semantic gap and improve target generalization. Practically, IDP offers a “free-lunch” style adaptation that minimizes additional data and computational burden during deployment while delivering robust cross-domain transfer in few-shot settings.

Abstract

Cross-Domain Few-Shot Learning (CDFSL) endeavors to transfer generalized knowledge from the source domain to target domains using only a minimal amount of training data, which faces a triplet of learning challenges in the meantime, i.e., semantic disjoint, large domain discrepancy, and data scarcity. Different from predominant CDFSL works focused on generalized representations, we make novel attempts to construct Intermediate Domain Proxies (IDP) with source feature embeddings as the codebook and reconstruct the target domain feature with this learned codebook. We then conduct an empirical study to explore the intrinsic attributes from perspectives of visual styles and semantic contents in intermediate domain proxies. Reaping benefits from these attributes of intermediate domains, we develop a fast domain alignment method to use these proxies as learning guidance for target domain feature transformation. With the collaborative learning of intermediate domain reconstruction and target feature transformation, our proposed model is able to surpass the state-of-the-art models by a margin on 8 cross-domain few-shot learning benchmarks.

Figures (11)

• Figure 1: Illustration of intermediate proxy reconstruction. a) Two representative sub-domains for reconstruction bases. b) Reconstructed intermediate proxies using the sources in a).
• Figure 2: Quantitative evaluations of content and style differences among source domains $\mathcal{S}$, target domains $\mathcal{T}$, and intermediate domain proxies $\mathcal{P}$. The content distance is calculated by VGG perception and Style distance is calculated by Gram distance using samples in mini-ImageNet (Detailed in Appendix A).
• Figure 3: Illustration of the proposed method. We first collect dense prototypes from source domains and use them to construct the intermediate proxy pool. Features in this pool are then employed to reconstruct the target domain, forming intermediate reconstructions. After that, the intermediate proxies are adopted as learning guidance for fast feature alignment of target and intermediate domains.
• Figure 4: Frequency distribution of randomly sampled distances between feature pairs from different domains (Higher values indicates better domain alignments.). a1) and a2): Domain-transferring ability of Baseline and Ours. b): Intermediate domain proxies boost knowledge-transferring capabilities.
• Figure 5: The effect of different size of class prototype $\bm{\mathrm{V}}_i^{t}$. All experiments are conducted under 5-way 5-shot conditions, and the vertical coordinates indicate the performance of our method.