Source-linked AI summary
Deep Visual Domain Adaptation
Gabriela Csurka
TL;DR
Visual domain adaptation must transfer knowledge from labeled source data to unlabeled target data despite domain mismatch and the high data demands of deep learning. This paper surveys deep architectures and adaptation trends, along with complementary improvement strategies, showing how they address challenging visual shifts across applications.
Problem
Deep models need extensive labeled data, while target domains are often unlabeled and differ in distribution from related labeled source domains.
Method
The paper compares four deep-adaptation strategies, surveys deep domain-adaptation model families, and reviews orthogonal techniques such as pseudo-labeling and entropy minimization.
Results
Deep representations provide stronger baselines than handcrafted features, while combining source fine-tuning with shallow adaptation can approach dedicated deep adaptation models.
Takeaways & Limitations
Deep visual domain adaptation spans feature reuse, fine-tuning, joint alignment, style transfer, and target-prediction-based improvements across several computer-vision tasks.
Abstract
from arXiv · showhide
Domain adaptation (DA) aims at improving the performance of a model on target domains by transferring the knowledge contained in different but related source domains. With recent advances in deep learning models which are extremely data hungry, the interest for visual DA has significantly increased in the last decade and the number of related work in the field exploded. The aim of this paper, therefore, is to give a comprehensive overview of deep domain adaptation methods for computer vision applications. First, we detail and compared different possible ways of exploiting deep architectures for domain adaptation. Then, we propose an overview of recent trends in deep visual DA. Finally, we mention a few improvement strategies, orthogonal to these methods, that can be applied to these models. While we mainly focus on image classification, we give pointers to papers that extend these ideas for other applications such as semantic segmentation, object detection, person re-identifications, and others.
I. INTRODUCTION
Deep visual domain adaptation addresses the need to transfer knowledge from labeled source data to unlabeled target data despite distribution mismatch. The paper motivates deep representations and surveys strategies for adapting them to challenging domain shifts.
- Domain adaptation transfers knowledge from a labeled source domain to an unlabeled target domain when domains differ but share a task.
- Deep learning improves visual recognition but requires large amounts of well-annotated data, making unlabeled target data valuable.
- Deep representations are more transferable because they encode category-level information, reduce domain bias, and disentangle factors of variation.
- Deep-network features provide stronger domain-adaptation baselines than handcrafted bag-of-visual-words representations, even without target adaptation.
- Directly using source-trained deep models works especially when domain shift is moderate, but challenging shifts such as images-to-sketches still require adaptation.
- The paper compares ways to exploit deep architectures, surveys recent deep visual adaptation trends, and discusses orthogonal improvement strategies.
II. DEEP LEARNING STRATEGIES
The paper organizes deep-learning strategies for domain adaptation from deep-feature baselines to fine-tuned and end-to-end models. Combining source fine-tuning with shallow adaptation can approach a dedicated deep adaptation network.
- Strategy taxonomy: Deep domain-adaptation strategies comprise shallow methods using deep features, source-trained deep architectures, fine-tuned deep features with shallow adaptation, and deep adaptation models.
- Shallow methods using deep features: Using deep features in traditional adaptation methods further improves target performance beyond a classifier trained on pre-trained representations.
- Training deep architectures on the source: Fine-tuning a deep network on the source can outperform feature-based baselines and shallow adaptation, but large domain differences may cause source overfitting.
- Shallow methods using fine-tuned deep features: Combining source fine-tuning with shallow distribution adaptation uses fine-tuned representations while avoiding a network architecture tailored specifically for adaptation.
- Comparison: On LandMarkDA, FT+SDAN performs close to DDAN, while DDAN performs best among the compared strategies.
- Comparison: The combined shallow strategy yields results close to a deep end-to-end model using the same discrepancy and classification losses.
III. DEEP DA MODELS
Deep domain-adaptation models jointly learn representations, source classification, and domain alignment, extending classical fixed-feature pipelines. The survey covers discriminative alignment, adversarial confusion, task-conditional alignment, parameter specialization, reconstruction, and image-style transfer.
- Architecture: Classical methods adapt fixed representations, whereas deep models learn representations, source classification, and domain alignment jointly end to end.
- Discriminative models: Discriminative models use source and target streams with shared, partially shared, or separate weights, applying source classification and feature-alignment losses.
- Alignment mechanisms: Alignment minimizes distribution discrepancy or increases domain confusion through losses including MMD, CORAL, Wasserstein, GAN, and gradient reversal.
- Class-conditional distribution alignment: Class-conditional approaches incorporate class predictions or decision-boundary disagreement so alignment accounts for the task rather than only marginal distributions.
- Network parameter adaptation: Some methods adapt target-stream parameters or use multi-stream architectures, specializing processing for each domain instead of enforcing fully invariant features.
- Transfer domain style: Image-to-image transfer generates target-like source images that retain source labels, using paired styles or unpaired translation for appearance shifts.
IV. ORTHOGONAL IMPROVEMENT STRATEGIES
The survey describes auxiliary strategies that can be combined with deep adaptation architectures, including pseudo-labeling, curricula, entropy minimization, and self-ensembling. These methods use target predictions, confidence, or consistency to improve adaptation.
- Orthogonal machine-learning strategies can be combined with deep adaptation architectures to further improve performance.
- Pseudo-labeling: Pseudo-labeling treats sufficiently reliable predicted target labels as supervision, extending adaptation toward a semi-supervised setting.
- Pseudo-labeling: Pseudo-labeling supports adaptation beyond classification, including fashion-product matching and person re-identification.
- Curriculum learning: Curriculum learning introduces the most confident pseudo-labeled target samples first and adds less confident samples later to reduce noise.
- Curriculum learning: Curriculum-based adaptation has also been applied to cross-domain person re-identification and image segmentation.
- Conditional entropy minimization: Conditional entropy minimization improves target-domain decision boundaries, with adversarial variants and consistency-based formulations extending the strategy.
- Self-ensemble learning: Self-ensemble learning enforces consistent target predictions under perturbations such as augmentation, dropout, and noise.