Generalized Meta-FDMixup: Cross-Domain Few-Shot Learning Guided by Labeled Target Data

Yuqian Fu, Yanwei Fu, Jingjing Chen, Yu-Gang Jiang

Abstract

Cross-Domain Few-Shot Learning (CD-FSL) addresses the challenge of generalizing models to target domains with significant visual shifts from the source training data. This paper introduces meta-FDMixup, a novel framework that leverages a minimal amount of labeled target data to guide the meta-learning process. By integrating cross-domain image mixup and feature disentanglement, the proposed method effectively bridges the domain gap, significantly improving performance on diverse target datasets while maintaining source domain accuracy.

Introduction

Standard Few-Shot Learning (FSL) trains models to recognize novel classes using only a handful of examples. However, standard FSL assumes that training (source) and testing (target) data come from the same domain. When there is a massive visual shift—such as training on general objects and testing on fine-grained bird species or medical images—standard FSL models fail catastrophically. This challenge is known as Cross-Domain Few-Shot Learning (CD-FSL).

Comparison of Few-shot Learning, Domain Adaptation, and Cross-domain Few-shot Learning paradigms.

Existing CD-FSL methods typically rely solely on source data or large volumes of unlabeled target data. This paper argues that in real-world scenarios, labeling a tiny amount of target data (e.g., 5 images per class) is often more practical than curating massive unlabeled datasets. The authors explore how to effectively inject this extremely sparse "auxiliary" target data into the meta-learning process to rescue model performance.

The Meta-FDMixup Architecture

To leverage a tiny auxiliary dataset without overfitting, the authors introduce the meta-FDMixup network. The architecture consists of two primary stages: pre-training on the source dataset and meta-training using both source and auxiliary target data.

The overall architecture of meta-FDMixup, illustrating the pre-training (left) and meta-training (right) phases.

1. Cross-Domain Mixup Module

During the meta-training phase, the model samples "episodes" from both the source dataset and the auxiliary target dataset. It then mathematically mixes the query images of these two domains using a linear combination (Mixup), while keeping the support (reference) sets unchanged. This process artificially inflates the diversity and volume of the auxiliary data. The mixup ratio $\lambda$ serves as a ground-truth confidence score for the FSL classification loss, penalizing the model if it fails to proportionally attribute the mixed image to both source and target support sets.

2. Feature Disentanglement

A specialized neural module splits extracted image features into two distinct streams:

  • Domain-irrelevant features: Used by the FSL classifier to identify the class of the object, regardless of its domain.
  • Domain-specific features: Used by a secondary "domain classifier" to predict whether the image originated from the source or target dataset.

By forcing the network to separate class-relevant information from domain-specific noise, the model learns robust representations that transfer more effectively across the domain gap.

Experimental Evaluation

The model was evaluated using mini-ImageNet as the source dataset and several fine-grained datasets as targets: CUB (birds), Cars, Places, and Plantae. To maintain the integrity of the FSL setup, auxiliary data was sampled from the target dataset's base classes, while evaluation was performed on entirely disjoint novel classes.

Key Results

  • Performance Gains: On the 5-way 5-shot task, meta-FDMixup achieved significant accuracy: 79.46% on CUB, 66.52% on Cars, 78.92% on Places, and 69.22% on Plantae.
  • Baseline Comparison: Compared to a standard Graph Neural Network (GNN) baseline, meta-FDMixup improved accuracy by ~17% absolute on the CUB dataset (62.25% to 79.46%).
  • Source Preservation: Training with auxiliary target data did not cause "catastrophic forgetting"; instead, it actually improved the model's accuracy on the original source dataset.

Insights and Limitations

A key finding of the study is that the timing of target data introduction is critical; introducing auxiliary data during pre-training can be detrimental, whereas introduction during meta-training provides a strong inductive bias. However, the authors noted a "law of diminishing marginal utility," where performance gains flatten quickly after the first 5–15 labeled images per class.

While the current pipeline requires training a specific model for each target domain, the results validate the practicality of using extremely few labeled target images to bridge the gap in cross-domain scenarios, encouraging a shift toward this realistic hybrid setting in CD-FSL research.

Summary updated

BibTeX
@article{fu2022generalized,
  author = {Yuqian Fu and Yanwei Fu and Jingjing Chen and Yu-Gang Jiang},
  title = {{Generalized Meta-FDMixup: Cross-Domain Few-Shot Learning Guided by Labeled Target Data}},
  journal = {IEEE Trans. Image Process.},
  year = {2022},
  doi = {10.1109/tip.2022.3219237},
}
Figure