PaperlayerView exact revision on arXiv

Source-linked AI summary

Heterogeneous Domain Generalization via Domain Mixup

Yufei Wang, Haoliang Li, Alex C. Kot

arXiv:2009.05448v1cs.CV

TL;DR

The central question is how to generalize DCNN features across domains when source distributions differ and target domains have fewer labeled samples. The paper applies two domain-mixup strategies and reports that its method outperforms state-of-the-art baselines on Visual Decathlon.

  • Problem

    The central question is how to generalize DCNN features across domains when source distributions differ and target domains have fewer labeled samples.

    Source: §1 · p. 1, §2.1 · p. 2

  • Method

    The method regularizes the model using convex combinations of samples from different domains through two strategies: mixing between training examples or among multiple domains.

    Source: §1 · p. 2

  • Results

    On Visual Decathlon, the authors report that their method outperforms state-of-the-art baselines on classification accuracy and VD-score.

    Source: §4 · p. 5, §3 · p. 5

  • Takeaways & Limitations

    The authors describe the resulting feature extractor as more robust and helpful when used off-the-shelf.

    Source: §4 · p. 5

    The HDG setup assumes that label spaces are disjoint across domains.

    Source: §2.1 · p. 2

Abstract

from arXiv · show

One of the main drawbacks of deep Convolutional Neural Networks (DCNN) is that they lack generalization capability. In this work, we focus on the problem of heterogeneous domain generalization which aims to improve the generalization capability across different tasks, which is, how to learn a DCNN model with multiple domain data such that the trained feature extractor can be generalized to supporting recognition of novel categories in a novel target domain. To solve this problem, we propose a novel heterogeneous domain generalization method by mixing up samples across multiple source domains with two different sampling strategies. Our experimental results based on the Visual Decathlon benchmark demonstrates the effectiveness of our proposed method. The code is released in \url{https://github.com/wyf0912/MIXALL}

1. INTRODUCTION

The paper addresses heterogeneous domain generalization, where source and target domains have disjoint label spaces, with two mixup algorithms that combine samples across domains. The authors report that their method outperforms state-of-the-art baselines on the Visual Decathlon benchmark.

1. INTRODUCTION

  • Pretraining followed by fine-tuning or feature extraction may still perform poorly when training and test domains differ, limiting classifier generalization.These approaches are presented as responses to limited large-scale training data; the domain gap can undermine their practical performance.Source: §1 · p. 1
  • Domain adaptation assumes access to unlabeled or sparsely labeled target samples during training, while UDA, PDA, and OSDA address differing label-space conditions.UDA distinguishes common and private labels, PDA treats target labels as a subset of source labels, and OSDA treats private labels as an “unknown” class.Source: §1 · p. 1
  • Heterogeneous domain generalization assumes disjoint source and target label spaces, unlike domain adaptation and domain generalization, which require shared label spaces.Domain generalization trains only on source-domain data because the target domain is unavailable during training [10] [12].Source: §1 · p. 1
  • A prior HDG approach uses a feature-critic network with meta-learning regularization and splits training data into meta-training and meta-validation sets to simulate domain shift.Source: §1 · p. 2
  • Two mixup algorithms regularize the model using convex combinations of samples across domains, modeling either linear behavior between training examples or relationships among multiple domains.The method is built on the mixup strategy.Source: §1 · p. 2
  • The authors report that their method outperforms state-of-the-art baselines for heterogeneous domain generalization on the Visual Decathlon benchmark.Source: §1 · p. 2

2. PROPOSED METHOD

The method trains a shared feature extractor on labeled source domains, then adapts it to target domains with fewer labeled samples. It extends mixup to combine data across source domains with pairwise or all-domain sampling.

2.1. Problem Definition and Notation

  • HDG learns a feature extractor shared across tasks despite disjoint domain label spaces, training on source domains and evaluating target-domain classifiers using extracted features.Target domains have fewer labeled samples; a shallow classifier such as an SVM or KNN is trained on their extracted features.Source: §2.1 · p. 2

2.2. A baseline for HDG

  • The aggregation baseline trains a feature extractor, freezes it to extract target-domain features, and trains classifiers using target-domain training data.The classifiers are tested on the target domain’s testing set.Source: §2.2 · p. 2

2.3. MIXUP for HDG

  • HDG mixup integrates information across source domains, addressing their differing data distributions and disjoint label spaces.The MIXUP strategy mixes samples from two arbitrary domains per iteration, while MIX-ALL mixes samples from all source domains; β controls the mixing weights, and increasing β makes generated samples more definite and less random.Source: §2.3 · p. 3

3. EXPERIMENT

Experiments on Visual Domain Decathlon report that the proposed method outperforms baselines with SVM and KNN classifiers; ablations show performance depends on mixup strength.

Results

  • On VD, the method significantly outperforms the listed baselines in SVM classification accuracy and VD Score, and achieves much better performance with KNN.The baselines include MetaReg, CrossGrad, Reptile, and Feature-Critic Networks.Source: Table 1 · p. 4, Table 2 · p. 4, §3 · p. 5

Dataset

  • The VD benchmark has ten heterogeneous domains with disjoint label spaces; six larger domains are sources and four smaller domains are targets.All images were resized to a shorter side of 72 pixels.Source: §3 · p. 4

Networks and Training Details

  • Evaluation uses target-domain features to train SVM and KNN classifiers, with grid-search tuning and cross-validation on each target domain’s training split.Experiments use ResNet-18, freeze its first four blocks, initialize from AGG weights, and train end-to-end with AMSGrad for 20k iterations; mixup settings are α = 0.4 and β = 6.Source: §3 · p. 4

Ablation Study

  • For α, performance in accuracy and VD-score first improves and then drops; near zero, the model degenerates into the proposed baseline.The authors conjecture that the initial improvement reflects enhanced robustness and the decline reflects underfitting.Source: Figure 2 · p. 5, §3 · p. 5
  • For MIX-ALL, average accuracy drops at β = 0 or 1; small β may add too much noise, very large β returns to the baseline, and performance falls as n grows.The authors relate small β to low expectations of max(λ) and σ(λ)^2.Source: Figure 3 · p. 5, §3 · p. 5

4. CONCLUSION

The paper proposes heterogeneous domain generalization through domain mixup, using two-domain and multi-domain strategies to augment training data. The authors report that the resulting feature extractor is more robust and that their methods outperform existing state-of-the-art baselines in classification accuracy and VD-score.

  • Domain mixup uses two strategies—mixup between two domains and across multiple domains—to sample augmented training data and produce a more robust feature extractor for off-the-shelf use.Source: §4 · p. 5
  • The proposed methods outperform existing state-of-the-art baselines in classification accuracy and VD-score.Source: §4 · p. 5

Related papers

Loading 2009.05448v1…