Deep Visual Domain Adaptation: A Survey

Mei WangWeihong Deng

article2018Neurocomputing2,371 citations

Systematizes deep visual domain adaptation methods by divergence scenarios and loss formulations while examining their practical deployment across advanced computer vision tasks like semantic segmentation and object detection.

Listen

Modern computer vision systems rely heavily on large collections of labeled data, but collecting and annotating new datasets for every operational deployment is expensive and time-consuming. While vast amounts of labeled imagery already exist, machine learning models experience severe performance degradation when applied to new environments due to differences in lighting, camera angles, image quality, and other environmental variations. This underlying distribution mismatch prevents organizations from reliably deploying models trained on auxiliary data directly to new targets. To address this issue, the article evaluates and synthesizes deep visual domain adaptation methods, which embed adaptation criteria directly into deep neural network training to produce transferable, domain-invariant representations.

The article conducts a comprehensive review and comparative analysis of deep visual adaptation techniques across diverse computer vision tasks. It establishes a taxonomy based on how data domains diverge and classifies the core technical approaches into three primary methodological paradigms: discrepancy-based methods that minimize statistical or architectural differences, adversarial approaches that use discriminator models to confuse domain origins, and reconstruction-based strategies that reconstruct source and target samples to preserve essential shared characteristics. The review evaluates these strategies across established benchmark datasets and practical computer vision use cases, including image classification, face recognition, object detection, and semantic segmentation.

The findings show that applying adaptation mechanisms directly within deep neural network pipelines consistently outperforms standard non-adapted deep architectures. On benchmark transfer tasks, adaptation models substantially improved classification accuracy, frequently raising performance by 10 to over 20 percentage points compared to standard baseline networks. Furthermore, unsupervised adaptation has emerged as the most practical paradigm because models trained on very small sets of labeled target data tend to overfit to source distributions. For more complex shifts or distinct domains, multi-step adaptation utilizing intermediate representations provides a viable bridge to transfer knowledge where single-step transfer fails.

These results demonstrate that organizations can significantly reduce data labeling costs and accelerate deployment schedules by adapting existing models rather than building new annotated datasets from scratch. Decision-makers evaluating computer vision systems should focus engineering investments on unsupervised adversarial or discrepancy-matching architectures when transitioning models into new visual environments. Future implementation efforts should explore heterogeneous domain adaptation and address scenarios where target environments do not share the exact same categories as the training source data.

The current body of research relies primarily on homogeneous settings where the source and target feature dimensions match, leaving heterogeneous cross-media applications less developed. Readers should recognize that empirical performance varies depending on how closely the source and target environments are aligned, meaning that complex visual deployments may require additional pilot validation before operational release.

Beijing University of Posts and Telecommunications School of Information and Communication Engineering computer vision computer vision applications deep learning Domain Deep Target

arXiv: 1802.03601

Book: Domain-Adversarial Training of Neural Networks, Yaroslav Ganin et al. (2016). Reading this foundational work on domain-adversarial neural networks provides the essential theoretical and architectural grounding needed to understand the adversarial adaptation techniques surveyed in the source.
Paper: Deep Domain Confusion: Maximizing for Domain Invariance, Eric Tzeng et al. (2014). Understanding this early breakthrough in embedding domain confusion within deep architectures is vital for following the historical progression and taxonomy detailed in the survey.
Paper: Learning Transferable Features with Deep Adaptation Networks, Mingsheng Long et al. (2015). Familiarity with deep adaptation networks and maximum mean discrepancy regularization is necessary to comprehend the discrepancy-reduction methods categorized in the source paper.
Paper: Adversarial Discriminative Domain Adaptation, Eric Tzeng et al. (2017). Exploring adversarial discriminative domain adaptation offers critical insight into untied weight encoders and adversarial loss variants that heavily influence modern transfer learning.
Paper: A theory of learning from different domains, Shai Ben-David et al. (2010). Reviewing this foundational learning theory establishes the formal generalization bounds and divergence measures that underpin the theoretical analysis of domain adaptation discussed in the survey.

Paper: CyCADA: Cycle-Consistent Adversarial Domain Adaptation, Judy Hoffman et al. (2018). This paper extends the survey's coverage of domain adaptation by introducing cycle-consistent adversarial training to bridge pixel-level and feature-level visual domains.
Paper: Moment Matching for Multi-Source Domain Adaptation, Xingchao Peng et al. (2018). Building directly upon the single-source adaptation methods reviewed in the survey, this work generalizes the framework to handle multi-source domain adaptation across large-scale benchmarks.
Paper: Deep Clustering for Unsupervised Learning of Visual Features, Mathilde Caron et al. (2018). This study continues the exploration of unsupervised visual representations by applying scalable deep clustering techniques that bypass the need for explicit domain labels.
Paper: Transformers in Vision: A Survey, Salman Khan et al. (2021). Extending beyond the convolutional architectures featured prominently in the survey, this work examines how emerging Vision Transformer models revolutionize visual adaptation and representation.
Paper: Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation, Liang-Chieh Chen et al. (2018). This research applies advanced encoder-decoder structures and atrous convolutions to semantic image segmentation, advancing the computer vision applications highlighted in the survey.

Cover for Deep Visual Domain Adaptation: A Survey

Abstract

Deep domain adaption has emerged as a new learning technique to address the lack of massive amounts of labeled data. Compared to conventional methods, which learn shared feature subspaces or reuse important source instances with shallow representations, deep domain adaption methods leverage deep networks to learn more transferable representations by embedding domain adaptation in the pipeline of deep learning. There have been comprehensive surveys for shallow domain adaption, but few timely reviews the emerging deep learning based methods. In this paper, we provide a comprehensive survey of deep domain adaptation methods for computer vision applications with four major contributions. First, we present a taxonomy of different deep domain adaption scenarios according to the properties of data that define how two domains are diverged. Second, we summarize deep domain adaption approaches into several categories based on training loss, and analyze and compare briefly the state-of-the-art methods under these categories. Third, we overview the computer vision applications that go beyond image classification, such as face recognition, semantic segmentation and object detection. Fourth, some potential deficiencies of current methods and several future directions are highlighted.

I INTRODUCTION
II Overview
II-A Notations and Definitions
II-B Different Settings of Domain Adaptation
III Approaches of Deep Domain Adaptation
III-A Categorization of One-Step Domain Adaptation
III-B Categorization of Multi-Step Domain Adaptation
IV ONE-STEP DOMAIN ADAPTATION
IV-A Homogeneous Domain Adaptation
IV-A1 Discrepancy-Based Approaches
IV-A2 Adversarial-Based Approaches
IV-A3 Reconstruction-Based Approaches
IV-A4 Hybrid Approaches
IV-B Heterogeneous Domain Adaptation
IV-B1 Discrepancy-Based Approach
IV-B2 Adversarial-Based Approach
IV-B3 Reconstruction-Based Approach
V Multi-Step Domain Adaptation
V-A Hand-Crafted Approaches
V-B Instance-Based Approaches
V-C Representation-Based Approaches
VI Application of Deep Domain Adaptation
VI-A Image Classification
VI-B Face Recognition
VI-C Object Detection
VI-D Semantic Segmentation
VI-E Image-to-Image Translation
VI-F Person Re-identification
VI-G Image Captioning
VII Conclusion
VIII Acknowledgements
References

Knowls

Knowl 1 — Taxonomy and Formal Problem Settings of Visual Domain Adaptation
definition
In visual domain adaptation (DA), a domain $\mathcal{D}$ is defined by a feature space $\mathcal{X}$ and a marginal probability distribution $P(X)$ , where $X = \{x_1, \dots, x_n\} \in \mathcal{X}$ . A task $\mathcal{T}$ consists of a label space $\mathcal{Y}$ and an objective predictive function $f(\cdot)$ , interpreted probabilistically as the conditional probability distribution $P(Y|X)$ . The learning problem involves a source domain $\mathcal{D}^s = \{\mathcal{X}^s, P(X^s)\}$ with task $\mathcal{T}^s = \{\mathcal{Y}^s, P(Y^s|X^s)\}$ possessing abundant labeled data $\mathcal{D}^s = \{(x_i^s, y_i^s)\}$ , and a target domain $\mathcal{D}^t = \{\mathcal{X}^t, P(X^t)\}$ with task $\mathcal{T}^t = \{\mathcal{Y}^t, P(Y^t|X^t)\}$ consisting of partially labeled data $\mathcal{D}^{tl}$ and unlabeled data $\mathcal{D}^{tu}$ ( $\{x_i^t\}$ ). Domain adaptation operates under the transductive assumption that the source and target tasks are identical ( $\mathcal{T}^s = \mathcal{T}^t$ ), while the domains diverge ( $\mathcal{D}^s \neq \mathcal{D}^t$ ).

Domain adaptation problems are categorized according to three foundational axes:
1. Feature Space Divergence:
  - Homogeneous DA: Feature spaces are identical ( $\mathcal{X}^s = \mathcal{X}^t$ ) with equal dimensionality ( $d^s = d^t$ ), but marginal data distributions differ ( $P(X^s) \neq P(X^t)$ ).
  - Heterogeneous DA: Feature spaces are non-equivalent ( $\mathcal{X}^s \neq \mathcal{X}^t$ ) and feature dimensions may differ ( $d^s \neq d^t$ ), occurring across different image modalities (e.g., RGB vs. depth, photos vs. sketches) or cross-media data (e.g., text vs. image).
2. Target Label Availability:
  - Supervised DA: A small, insufficient set of labeled target instances $\mathcal{D}^{tl}$ is available during training.
  - Semi-Supervised DA: Both limited labeled target instances $\mathcal{D}^{tl}$ and redundant unlabeled target instances $\mathcal{D}^{tu}$ are available.
  - Unsupervised DA: No labeled target data is available ( $\{x_i^t\} \in \mathcal{D}^{tu}$ only).
3. Domain Relatedness Distance:
  - One-Step DA: Knowledge is transferred directly between source and target domains assuming substantial domain overlap.
  - Multi-Step (Transitive) DA: The source and target domains share little direct overlap, requiring one or more intermediate domains to construct a smooth bridging path.
Knowl 2 — Taxonomy of One-Step Deep Domain Adaptation Approaches
model/method
Deep domain adaptation integrates domain discrepancy reduction into the deep representation learning pipeline via back-propagation to obtain representations that are simultaneously discriminative for the primary task and invariant across domains. One-step deep domain adaptation approaches are categorized into three principal paradigms based on their training objectives:
1. Discrepancy-Based Approaches: Minimize domain divergence by fine-tuning deep networks on target data using explicit criteria:
  - Class Criterion: Uses target labels, soft labels, pseudo-labels, or semantic attributes to guide domain alignment.
  - Statistic Criterion: Aligns statistical distribution moments between source and target representations in latent layers (e.g., via Maximum Mean Discrepancy, covariance alignment, or central moment discrepancy).
  - Architecture Criterion: Modifies network architecture components (such as weight regularization across streams, domain-specific batch normalization, or domain-guided dropout) to learn transferable features.
  - Geometric Criterion: Constructs intermediate subspaces on Grassmannian geodesic paths connecting source and target manifolds.
2. Adversarial-Based Approaches: Employ domain discriminators in a minimax game to encourage domain confusion:
  - Generative Models: Integrate generative adversarial networks (GANs) conditioned on noise or source images to synthesize labeled target-like images.
  - Non-Generative Models: Map source and target data into a shared invariant feature space using domain-confusion or gradient reversal losses without synthetic image generation.
3. Reconstruction-Based Approaches: Utilize auxiliary data reconstruction to preserve domain-specific characteristics while enforcing shared representation invariance:
  - Encoder-Decoder Reconstruction: Combine encoders with decoders to reconstruct target inputs or partition representations into shared and private components.
  - Adversarial Reconstruction: Measure cyclic reconstruction errors across domain mappings using dual GAN discriminators.
Knowl 3 — Statistical Distribution Alignment in Deep Domain Adaptation
model/method
Statistic-criterion discrepancy methods align the latent feature distributions $P(X^s)$ and $P(X^t)$ across deep network layers. Several statistical metrics are utilized:
1. Maximum Mean Discrepancy (MMD): Computes the distance between empirical kernel mean embeddings in a Reproducing Kernel Hilbert Space (RKHS) $\mathcal{H}$ associated with feature map $\phi(\cdot)$ :
$\text{MMD}^2(\mathcal{D}^s, \mathcal{D}^t) = \left\| \frac{1}{M}\sum_{i=1}^M \phi(x_i^s) - \frac{1}{N}\sum_{j=1}^N \phi(x_j^t) \right\|_\mathcal{H}^2$

Deep Domain Confusion (DDC) optimizes cross-entropy classification loss $\mathcal{L}_C(X_L, y)$ alongside an MMD penalty on an adaptation layer:

$\mathcal{L} = \mathcal{L}_C(X_L, y) + \lambda \text{MMD}^2(X^s, X^t)$

Extensions include Deep Adaptation Networks (DAN), which apply multi-kernel MMD across multiple layers; Joint Adaptation Networks (JAN), which align joint distributions of input features and output labels via Joint MMD (JMMD); and Weighted MMD, which corrects for class weight bias between domains.
1. Correlation Alignment (Deep CORAL): Matches second-order statistics (covariance matrices $C_S$ and $C_T$ ) of source and target activations in deep feature spaces with dimension $d$ :
$\mathcal{L}_{\text{CORAL}} = \frac{1}{4d^2} \|C_S - C_T\|_F^2$

where $\|\cdot\|_F^2$ denotes the squared matrix Frobenius norm.
1. Central Moment Discrepancy (CMD): Matches higher-order sample central moments of distributions over compact intervals $[a, b]^N$ up to order $K$ :
$\text{CMD}_K(X^s, X^t) = \frac{1}{|b - a|} \|E(X^s) - E(X^t)\|_2 + \sum_{k=2}^K \frac{1}{|b - a|^k} \|C_k(X^s) - C_k(X^t)\|_2$

where $E(X) = \frac{1}{|X|} \sum_{x \in X} x$ is the empirical expectation and $C_k(X) = E((x - E(X))^k)$ is the vector of all $k$ -th order central moments.
1. Association Loss ( $\mathcal{L}_{\text{assoc}}$ ): Enforces statistical associations by constraining two-step round-trip transition probabilities $P_{ij}^{aba}$ between source and target representations to approximate a uniform distribution over class labels.
Knowl 4 — Class, Architectural, and Geometric Alignment Criteria in Deep DA
model/method
Discrepancy-based deep domain adaptation leverages non-statistical criteria to guide cross-domain alignment:
1. Class Criteria:
  - Soft Labels: Modifies softmax logits $z_i$ with a temperature parameter $T$ via $q_i = \frac{\exp(z_i / T)}{\sum_j \exp(z_j / T)}$ to preserve inter-class geometric relationships across domains during fine-tuning.
  - Metric Learning: Employs semantic alignment and separation losses or marginal Fisher analysis combined with MMD to minimize intra-class compactness $S_c$ and maximize inter-class separability $S_b$ across domains:
$\min_W S_c^{(M)} - \alpha S_b^{(M)} + \beta D_{ts}^{(M)}(X^s, X^t) + \gamma \sum_{m=1}^M (\|W^{(m)}\|_F^2 + \|b^{(m)}\|_2^2)$
- Semantic Attributes: Maps classes to fixed-length binary attribute vectors $a^c = (a_1^c, \dots, a_m^c)$ to compute posterior class probabilities $p(y|x) = \frac{p(y)}{p(a^y)} \prod_{m=1}^M p(a_m^y|x)$ for zero/few-label target adaptation.
- Pseudo-Labeling: Assigns target pseudo-labels $\hat{y}_j^t = \arg\max_c p(y_j^t = c | x_j^t)$ based on model posterior confidence to enable conditional distribution alignment.
1. Architectural Criteria:
  - Related Weight Regularization: Couples two-stream network parameters $\theta_j^s$ and $\theta_j^t$ using exponential penalties $r_w(\theta_j^s, \theta_j^t) = \exp(\|\theta_j^s - \theta_j^t\|^2) - 1$ or linear transformations $r_w(\theta_j^s, \theta_j^t) = \exp(\|a_j \theta_j^s + b_j - \theta_j^t\|^2) - 1$ .
  - Adaptive Batch Normalization (AdaBN): Normalizes target domain activations independently by recomputing channel-wise statistics $\mu(X^t)$ and $\sigma(X^t)$ :
$\text{BN}(X^t) = \lambda \left( \frac{x - \mu(X^t)}{\sigma(X^t)} \right) + \beta$
- Domain-Guided Dropout: Selectively mutes domain-irrelevant neurons based on the loss function gain $s_i = \mathcal{L}(g(x)_{\setminus i}) - \mathcal{L}(g(x))$ , where $g(x)_{\setminus i}$ denotes activations with the $i$ -th neuron zeroed.
1. Geometric Criteria: Deep Learning by Interpolating Domains (DLID) constructs intermediate datasets sampled along a geodesic flow on the Grassmann manifold connecting source and target subspaces, training deep sparse feature extractors across the path.
Knowl 5 — Adversarial-Based Deep Domain Adaptation
model/method
Adversarial deep domain adaptation employs a domain discriminator $D$ alongside a feature extractor/generator in a minimax objective to produce domain-invariant feature spaces. Approaches are divided into generative and non-generative frameworks:
1. Generative Adversarial DA:
  - Coupled GAN (CoGAN): Employs a pair of GANs (one for source, one for target) with tied weights in the first layers of the generators and final layers of the discriminators to learn a shared representation space that synthesizes paired cross-domain images from common noise vectors.
  - Pixel-Level Transfer (e.g., PixelDA): Trains a generator $G$ , task classifier $T$ , and domain discriminator $D$ conditioned on source images and noise vectors by optimizing:
$\min_{G, T} \max_D V(D, G) = \alpha \mathcal{L}_d(D, G) + \beta \mathcal{L}_t(T, G) + \gamma \mathcal{L}_c(G)$

where $\mathcal{L}_d$ is the adversarial loss, $\mathcal{L}_t$ is the target classification loss, and $\mathcal{L}_c$ is a masked pairwise mean squared error content-similarity loss preserving source foregrounds.
1. Non-Generative Adversarial DA:
  - Domain-Adversarial Neural Network (DANN): Implements a Gradient Reversal Layer (GRL) between a shared feature extractor and a domain discriminator, maximizing domain confusion while minimizing source classification error.
  - Adversarial Discriminative Domain Adaptation (ADDA): Uses untied source mapping $M^s$ and target mapping $M^t$ , training the target representation via an unconstrained adversarial loss against discriminator $D$ :
$\min_{M^s, C} \mathcal{L}_{\text{cls}}(X^s, Y^s) = -\mathbb{E}_{(x^s, y^s)} \sum_{k=1}^K \mathbf{1}_{[k = y^s]} \log C(M^s(x^s))$

$\min_D \mathcal{L}_{\text{adv}D}(X^s, X^t, M^s, M^t) = -\mathbb{E}_{x^s}[\log D(M^s(x^s))] - \mathbb{E}_{x^t}[\log(1 - D(M^t(x^t)))]$

$\min_{M^s, M^t} \mathcal{L}_{\text{adv}M}(M^s, M^t) = -\mathbb{E}_{x^t}[\log D(M^t(x^t))]$
- Selective Adversarial Networks (SAN): Employs multiple class-wise discriminators with probability weighting to enable partial domain adaptation when target label space is a subset of source label space.
Knowl 6 — Reconstruction-Based Deep Domain Adaptation
model/method
Reconstruction-based deep DA uses an auxiliary reconstruction task to ensure that learned representations retain essential intra-domain data structures while achieving inter-domain invariance. Approaches follow two main paradigms:
1. Encoder-Decoder Reconstruction Models:
  - Deep Reconstruction-Classification Network (DRCN): Uses a shared CNN encoder $\theta_{\text{enc}}$ , a supervised classification branch $\theta_{\text{lab}}$ for source data, and a deconvolutional decoder $\theta_{\text{dec}}$ for unsupervised target reconstruction, optimizing:
$\min \lambda \mathcal{L}_c(\{\theta_{\text{enc}}, \theta_{\text{lab}}\}) + (1 - \lambda) \mathcal{L}_r(\{\theta_{\text{enc}}, \theta_{\text{dec}}\})$

where $\mathcal{L}_c$ is source cross-entropy classification loss, $\mathcal{L}_r = \|x - f_r(x)\|_2^2$ is squared reconstruction loss, and $\lambda \in [0, 1]$ balances the objectives.
- Domain Separation Networks (DSN): Decomposes representations into private subspaces ( $E_p^s, E_p^t$ ) and a shared subspace ( $E_c$ ) via soft orthogonality constraints, feeding concatenated private and shared embeddings into a shared decoder to reconstruct the original inputs.
- Marginalized Stacked Denoising Autoencoders (mSDA): Marginalizes artificial noise in closed form, enabling scalable linear reconstruction across domain features.
1. Adversarial Reconstruction Models:
  - Cycle-Consistent GANs (CycleGAN, DualGAN, DiscoGAN): Learn dual mappings $G: X \to Y$ and $F: Y \to X$ using two discriminators ( $D_X, D_Y$ ) with an adversarial loss $\mathcal{L}_{\text{GAN}}$ and an explicit $L_1$ cycle-consistency reconstruction loss:
$\mathcal{L}_{\text{cyc}}(G, F) = \mathbb{E}_{x \sim P(X)}[\|F(G(x)) - x\|_1] + \mathbb{E}_{y \sim P(Y)}[\|G(F(y)) - y\|_1]$

ensuring $F(G(X)) \approx X$ and $G(F(Y)) \approx Y$ for unpaired image domain translation.
Knowl 7 — Heterogeneous Deep Domain Adaptation Mechanisms
model/method
Heterogeneous deep domain adaptation addresses scenarios where the source and target feature spaces diverge ( $\mathcal{X}^s \neq \mathcal{X}^t$ ) and feature dimensions differ ( $d^s \neq d^t$ ). Methods are organized into two primary scenarios based on the nature of the cross-domain gap:
1. Cross-Modality Image Adaptation (RGB vs. Depth, Visual Light vs. Near-Infrared, Photos vs. Sketches):
  - Images across modalities can be resized to identical spatial dimensions, enabling discrepancy-based fine-tuning and mid-level feature adaptation.
  - Cross-Modal Distillation: Uses mid-level CNN representations from a paired modality (e.g., RGB) as a supervisory signal to train a target modality network (e.g., depth).
  - Multi-Modal Feature Fusion: Networks pre-trained on RGB data are fine-tuned on target modalities and fused at intermediate fully connected layers (e.g., fc6) for joint prediction.
  - Multi-Adversarial Synthesis: Employs cycle-consistent adversarial networks with multi-scale generator feature maps to translate between modalities (e.g., photo-sketch synthesis) without synthesis artifacts.
2. Cross-Media Adaptation (Text vs. Image, Language vs. Vision):
  - Features across disparate media cannot be directly aligned by spatial resizing, requiring shared semantic embedding architectures.
  - Weakly-Shared Deep Transfer Networks (DTN): Feeds paired text and image data into separate Stacked Autoencoders (SAEs) connected to weakly parameter-shared layers at the top.
  - Transfer Neural Trees (TNT): Utilizes two-stream deep networks to extract modality-specific representations, followed by a transfer neural decision forest with stochastic pruning to adapt predictive neurons.
  - Stacked Generative Adversarial Networks (StackGAN): Translates text descriptions into photo-realistic images via conditioning augmentation across two stages: Stage-I generates primitive shapes and basic colors at low resolution; Stage-II generates high-resolution details conditioned on text embeddings and Stage-I results.
Knowl 8 — Mechanisms for Multi-Step (Transitive) Deep Domain Adaptation
model/method
When source and target domains have minimal overlap, direct one-step adaptation fails. Multi-step (transitive or distant) domain adaptation connects distant domains via intermediate domains $\mathcal{D}^I$ that provide smooth bridge transitions. Intermediate domains are constructed and utilized via three selection mechanisms:
1. Hand-Crafted Selection: Intermediate domains are selected using human prior knowledge or external proxy tasks (e.g., utilizing nighttime light intensity satellite data as an intermediate bridge to adapt knowledge from daytime satellite imagery to regional economic poverty prediction).
2. Instance-Based Selection: Automatically selects relevant subsets of auxiliary data instances to build intermediate bridging domains. Distant Domain Transfer Learning (DDTL) selects auxiliary unlabeled instances by optimizing instance selection indicator vectors $v_S = (v_1^S, \dots, v_{n_S}^S)^T$ and $v_I = (v_1^I, \dots, v_{n_I}^I)^T$ ( $v_i \in \{0, 1\}$ ) alongside autoencoder reconstruction errors across source, intermediate, and target samples:
$J_1(f_e, f_d, v_S, v_T) = \frac{1}{n_S} \sum_{i=1}^{n_S} v_i^S \|\hat{x}_i^S - x_i^S\|_2^2 + \frac{1}{n_I} \sum_{i=1}^{n_I} v_i^I \|\hat{x}_i^I - x_i^I\|_2^2 + \frac{1}{n_T} \sum_{i=1}^{n_T} \|\hat{x}_i^T - x_i^T\|_2^2 + R(v_S, v_T)$

where $\hat{x}$ are autoencoder reconstructions parameterized by encoder $f_e$ and decoder $f_d$ , and $R(v_S, v_T)$ regularizes against trivial zero solutions. Unrelated source instances are eliminated as intermediate representations converge toward the target.
1. Representation-Based Selection: Progressive Neural Networks freeze previously trained network parameters on source and intermediate domains and establish lateral layer-wise connections to newly allocated target network streams, preventing catastrophic forgetting while transferring sequential domain representations.

Knowl 9 — Empirical Classification Accuracy of Deep DA vs. Non-Adaptation Baselines

data/table

Standard deep domain adaptation methods demonstrate consistent performance improvements over non-adapted deep baselines across standard visual benchmarks (Office-31 under AlexNet backbone, and Digits under VGG-16 backbone). Office-31 contains 31 classes across three domains: Amazon (A, 2817 images), DSLR (D, 498 images), and Webcam (W, 795 images).

Dataset / Reference	Source $\to$ Target	Baseline	Deep Domain Adaptation Methods
Office-31 ACC (%)		AlexNet	DDC	DAN	RTN	JAN	DANN
A $\to$ W		61.6 $\pm$ 0.5	61.8 $\pm$ 0.4	68.5	73.3 $\pm$ 0.3	75.2 $\pm$ 0.4	73.0 $\pm$ 0.5
D $\to$ W		95.4 $\pm$ 0.3	95.0 $\pm$ 0.5	96.0 $\pm$ 0.3	96.8 $\pm$ 0.2	96.6 $\pm$ 0.2	96.4 $\pm$ 0.3
W $\to$ D		99.0 $\pm$ 0.2	98.5 $\pm$ 0.4	99.0 $\pm$ 0.3	99.6 $\pm$ 0.1	99.6 $\pm$ 0.1	99.2 $\pm$ 0.3
A $\to$ D		63.8 $\pm$ 0.5	64.4 $\pm$ 0.3	67.0 $\pm$ 0.4	71.0 $\pm$ 0.2	72.8 $\pm$ 0.3	72.3 $\pm$ 0.3
D $\to$ A		51.1 $\pm$ 0.6	52.1 $\pm$ 0.6	54.0 $\pm$ 0.5	50.5 $\pm$ 0.3	57.5 $\pm$ 0.2	53.4 $\pm$ 0.4
W $\to$ A		49.8 $\pm$ 0.4	52.2 $\pm$ 0.4	53.1 $\pm$ 0.5	51.0 $\pm$ 0.1	56.3 $\pm$ 0.2	51.2 $\pm$ 0.5
Average		70.1	70.6	72.9	73.7	76.3	74.3
Office-31 ACC (%)		AlexNet	Deep CORAL	CMD	DLID	AdaBN	DANN
A $\to$ W		61.6	66.4	77.0 $\pm$ 0.6	51.9	74.2	73.0
D $\to$ W		95.4	95.7	96.3 $\pm$ 0.4	78.2	95.7	96.4
W $\to$ D		99.0	99.2	99.2 $\pm$ 0.2	89.9	99.8	99.2
A $\to$ D		63.8	66.8	79.6 $\pm$ 0.6	-	73.1	-
D $\to$ A		51.1	52.8	63.8 $\pm$ 0.7	-	59.8	-
W $\to$ A		49.8	51.5	63.3 $\pm$ 0.6	-	57.4	-
Average		70.1	72.1	79.9	-	76.7	-
Digits ACC (%)		VGG-16	DANN	CoGAN	ADDA
MNIST $\to$ USPS		75.2 $\pm$ 1.6	77.1 $\pm$ 1.8	91.2 $\pm$ 0.8	89.4 $\pm$ 0.2
USPS $\to$ MNIST		57.1 $\pm$ 1.7	73.0 $\pm$ 2.0	89.1 $\pm$ 0.8	90.1 $\pm$ 0.8
SVHN $\to$ MNIST		60.1 $\pm$ 1.1	73.9	-	76.0 $\pm$ 1.8

The empirical results indicate:

Discrepancy-based statistical alignment (e.g., JAN achieving $76.3\%$ average and CMD achieving $79.9\%$ average) and adversarial domain alignment (e.g., DANN at $74.3\%$ average) significantly outperform the standard AlexNet baseline ( $70.1\%$ ) on Office-31.
On digit adaptation tasks, adversarial generative and discriminative methods (CoGAN and ADDA) achieve accuracy exceeding $89\%$ - $91\%$ on MNIST $\leftrightarrow$ USPS, substantially outperforming the unadapted VGG-16 baseline ( $75.2\%$ and $57.1\%$ ).

Knowl 10 — Current Limitations and Open Frontiers in Deep Domain Adaptation
limitation
Existing deep domain adaptation literature exhibits several prominent structural limitations and challenges:
1. Over-Reliance on Homogeneous Assumptions: The vast majority of deep domain adaptation models assume identical feature spaces across domains ( $\mathcal{X}^s = \mathcal{X}^t$ ). Heterogeneous deep DA methods capable of bridging mismatched feature representations and cross-media gaps (such as text-to-vision) remain significantly underdeveloped.
2. Under-Exploration of Structured and Dense Prediction Tasks: Deep domain adaptation has been predominantly formulated and evaluated on global visual categorization (image classification). Complex visual tasks—such as object detection, semantic segmentation, face recognition across large variation, and person re-identification—involve spatial localization and localized feature shifts, where deep DA methods with scarce or zero target annotations are less developed.
3. Assumption of Identical Label Spaces: Standard deep DA methods assume that the source and target domains share an identical category set ( $\mathcal{Y}^s = \mathcal{Y}^t$ ). In real-world visual deployments, target datasets often contain only a subset of source classes (partial domain adaptation) or novel unseen classes (open-set domain adaptation). Applying standard distribution alignment under mismatched label spaces results in severe negative transfer.

Coverage note — Specific implementation details of individual domain-adapted vision applications (such as LSDA for object detection, SSPP-DAN for face recognition, SPGAN for person re-ID, and show-adapt-and-tell for image captioning) were summarized within their corresponding overarching algorithmic categories and limitations rather than extracted as isolated standalone knowls, ensuring focus on the survey's central taxonomies, formal models, and experimental syntheses.

References

1.M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017.
2.Y. Bengio. Learning deep architectures for ai. Foundations and Trends in Machine Learning, 2(1):1–127, 2009.
3.K. M. Borgwardt, A. Gretton, M. J. Rasch, H.-P. Kriegel, B. Schölkopf, and A. J. Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49–e57, 2006.
4.K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D. Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. arXiv preprint arXiv:1612.05424, 2016.
5.K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan. Domain separation networks. In Advances in Neural Information Processing Systems, pages 343–351, 2016.
6.L. Bruzzone and M. Marconcini. Domain adaptation problems: A dasvm classification technique and a circular validation strategy. IEEE transactions on pattern analysis and machine intelligence, 32(5):770–787, 2010.
7.P. P. Busto and J. Gall. Open set domain adaptation. In The IEEE International Conference on Computer Vision (ICCV), volume 1, page 3, 2017.
8.Z. Cao, M. Long, J. Wang, and M. I. Jordan. Partial transfer learning with selective adversarial networks. arXiv preprint arXiv:1707.07901, 2017.
9.F. M. Carlucci, L. Porzi, B. Caputo, E. Ricci, and S. R. Bulo`. Autodial: Automatic domain alignment layers. In International Conference on Computer Vision, 2017.
10.M. Chen, Z. Xu, K. Weinberger, and F. Sha. Marginalized denoising autoencoders for domain adaptation. arXiv preprint arXiv:1206.4683, 2012.
11.T.-H. Chen, Y.-H. Liao, C.-Y. Chuang, W.-T. Hsu, J. Fu, and M. Sun. Show, adapt and tell: Adversarial training of cross-domain image captioner. In The IEEE International Conference on Computer Vision (ICCV), volume 2, 2017.
12.W.-Y. Chen, T.-M. H. Hsu, Y.-H. H. Tsai, Y.-C. F. Wang, and M.-S. Chen. Transfer neural trees for heterogeneous domain adaptation. In European Conference on Computer Vision, pages 399–414. Springer, 2016.
13.Y. Chen, W. Li, C. Sakaridis, D. Dai, and L. Van Gool. Domain adaptive faster r-cnn for object detection in the wild. arXiv preprint arXiv:1803.03243, 2018.
14.Y. Chen, W. Li, and L. Van Gool. Road: Reality oriented adaptation for semantic segmentation of urban scenes. arXiv preprint arXiv:1711.11556, 2017.
15.Y.-H. Chen, W.-Y. Chen, Y.-T. Chen, B.-C. Tsai, Y.-C. F. Wang, and M. Sun. No more discrimination: Cross city adaptation of road scene segmenters. arXiv preprint arXiv:1704.08509, 2017.
16.S. Chopra, S. Balakrishnan, and R. Gopalan. Dlid: Deep learning for domain adaptation by interpolating between domains. In ICML workshop on challenges in representation learning, volume 2, 2013.
17.B. Chu, V. Madhavan, O. Beijbom, J. Hoffman, and T. Darrell. Best practices for fine-tuning visual classifiers to new domains. In Computer Vision–ECCV 2016 Workshops, pages 435–442. Springer, 2016.
18.W.-S. Chu, F. De la Torre, and J. F. Cohn. Selective transfer machine for personalized facial action unit detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3515–3522, 2013.
19.G. Csurka. Domain adaptation for visual applications: A comprehensive survey. arXiv preprint arXiv:1702.05374, 2017.
20.O. Day and T. M. Khoshgoftaar. A survey on heterogeneous transfer learning. Journal of Big Data, 4(1):29, 2017.
21.W. Deng, L. Zheng, G. Kang, Y. Yang, Q. Ye, and J. Jiao. Image-image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification. arXiv preprint arXiv:1711.07027, 2017.
22.J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In International conference on machine learning, pages 647–655, 2014.
23.L. Duan, D. Xu, and I. Tsang. Learning with augmented features for heterogeneous domain adaptation. arXiv preprint arXiv:1206.4660, 2012.
24.D. Eigen, C. Puhrsch, and R. Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in neural information processing systems, pages 2366–2374, 2014.
25.Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning, pages 1180–1189, 2015.
26.Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17(59):1–35, 2016.
27.L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2414–2423, 2016.
28.W. Ge and Y. Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint fine-tuning. arXiv preprint arXiv:1702.08690, 2017.
29.T. Gebru, J. Hoffman, and L. Fei-Fei. Fine-grained recognition in the wild: A multi-task domain adaptation approach. arXiv preprint arXiv:1709.02476, 2017.
30.M. Gheisari and M. S. Baghshah. Unsupervised domain adaptation via representation learning and adaptive classifier learning. Neurocomputing, 165:300–311, 2015.
31.M. Ghifary, W. Bastiaan Kleijn, M. Zhang, and D. Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In Proceedings of the IEEE international conference on computer vision, pages 2551–2559, 2015.
32.M. Ghifary, W. B. Kleijn, and M. Zhang. Domain adaptive neural networks for object recognition. In Pacific Rim International Conference on Artificial Intelligence, pages 898–904. Springer, 2014.
33.M. Ghifary, W. B. Kleijn, M. Zhang, D. Balduzzi, and W. Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In European Conference on Computer Vision, pages 597–613. Springer, 2016.
34.R. Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440–1448, 2015.
35.R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014.
36.X. Glorot, A. Bordes, and Y. Bengio. Domain adaptation for large-scale sentiment classification: A deep learning approach. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 513–520, 2011.
37.B. Gong, K. Grauman, and F. Sha. Connecting the dots with landmarks: Discriminatively learning domain-invariant features for unsupervised domain adaptation. In International Conference on Machine Learning, pages 222–230, 2013.
38.B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic flow kernel for unsupervised domain adaptation. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 2066–2073. IEEE, 2012.
39.I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014.
40.R. Gopalan, R. Li, and R. Chellappa. Domain adaptation for object recognition: An unsupervised approach. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages 999–1006. IEEE, 2011.
41.S. Gupta, J. Hoffman, and J. Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2827–2836, 2016.
42.P. Haeusser, T. Frerix, A. Mordvintsev, and D. Cremers. Associative domain adaptation. In International Conference on Computer Vision (ICCV), volume 2, page 6, 2017.
43.D. He, Y. Xia, T. Qin, L. Wang, N. Yu, T. Liu, and W.-Y. Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems, pages 820–828, 2016.
44.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
45.G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
46.G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets. Neural computation, 18(7):1527–1554, 2006.
47.J. Hoffman, S. Guadarrama, E. S. Tzeng, R. Hu, J. Donahue, R. Girshick, T. Darrell, and K. Saenko. Lsda: Large scale detection through adaptation. In Advances in Neural Information Processing Systems, pages 3536–3544, 2014.
48.J. Hoffman, S. Gupta, J. Leong, S. Guadarrama, and T. Darrell. Crossmodal adaptation for rgb-d detection. In Robotics and Automation (ICRA), 2016 IEEE International Conference on, pages 5032–5039. IEEE, 2016.
49.J. Hoffman, E. Tzeng, J. Donahue, Y. Jia, K. Saenko, and T. Darrell. One-shot adaptation of supervised deep convolutional models. arXiv preprint arXiv:1312.6204, 2013.
50.J. Hoffman, D. Wang, F. Yu, and T. Darrell. Fcns in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016.
51.S. Hong, W. Im, J. Ryu, and H. S. Yang. Sspp-dan: Deep domain adaptation network for face recognition with single sample per person. arXiv preprint arXiv:1702.04069, 2017.
52.S. Hong, J. Oh, H. Lee, and B. Han. Learning transferrable knowledge for semantic segmentation with deep convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3204–3212, 2016.
53.J. Hu, J. Lu, and Y.-P. Tan. Deep transfer metric learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 325–333, 2015.
54.X. Huang and S. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. arXiv preprint arXiv:1703.06868, 2017.
55.N. Inoue, R. Furuta, T. Yamasaki, and K. Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. arXiv preprint arXiv:1803.11365, 2018.
56.S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pages 448–456, 2015.
57.P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. arXiv preprint arXiv:1611.07004, 2016.
58.M. Kan, S. Shan, and X. Chen. Bi-shifting auto-encoder for unsupervised domain adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3846–3854, 2015.
59.T. Kim, M. Cha, H. Kim, J. Lee, and J. Kim. Learning to discover crossdomain relations with generative adversarial networks. arXiv preprint arXiv:1703.05192, 2017.
60.A. Kolesnikov and C. H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In European Conference on Computer Vision, pages 695–711. Springer, 2016.
61.P. Kontschieder, M. Fiterau, A. Criminisi, and S. Rota Bulo. Deep neural decision forests. In Proceedings of the IEEE International Conference on Computer Vision, pages 1467–1475, 2015.
62.A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
63.B. Kulis, K. Saenko, and T. Darrell. What you saw is not what you get: Domain adaptation using asymmetric kernel transforms. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1785–1792. IEEE, 2011.
64.C. H. Lampert, H. Nickisch, and S. Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 951–958. IEEE, 2009.
65.C. Li and M. Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In European Conference on Computer Vision, pages 702–716. Springer, 2016.
66.D. Li, Y. Yang, Y.-Z. Song, and T. M. Hospedales. Deeper, broader and artier domain generalization. In Computer Vision (ICCV), 2017 IEEE International Conference on, pages 5543–5551. IEEE, 2017.
67.Y. Li, K. Swersky, and R. Zemel. Generative moment matching networks. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pages 1718–1727, 2015.
68.Y. Li, N. Wang, J. Liu, and X. Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017.
69.Y. Li, N. Wang, J. Shi, J. Liu, and X. Hou. Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779, 2016.
70.M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. In Advances in neural information processing systems, pages 469–477, 2016.
71.W. Liu, Z. Wang, X. Liu, N. Zeng, Y. Liu, and F. E. Alsaadi. A survey of deep neural network architectures and their applications. Neurocomputing, 234:11–26, 2017.
72.X. Liu, L. Song, X. Wu, and T. Tan. Transferring deep representation for nir-vis heterogeneous face recognition. In Biometrics (ICB), 2016 International Conference on, pages 1–8. IEEE, 2016.
73.M. Long, Y. Cao, J. Wang, and M. Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97–105, 2015.
74.M. Long, J. Wang, and M. I. Jordan. Deep transfer learning with joint adaptation networks. arXiv preprint arXiv:1605.06636, 2016.
75.M. Long, H. Zhu, J. Wang, and M. I. Jordan. Unsupervised domain adaptation with residual transfer networks. In Advances in Neural Information Processing Systems, pages 136–144, 2016.
76.H. Lu, L. Zhang, Z. Cao, W. Wei, K. Xian, C. Shen, and A. van den Hengel. When unsupervised domain adaptation meets tensor representations. In The IEEE International Conference on Computer Vision (ICCV), volume 2, 2017.
77.P. Mittal, M. Vatsa, and R. Singh. Composite sketch recognition via deep network-a transfer learning approach. In Biometrics (ICB), 2015 International Conference on, pages 251–256. IEEE, 2015.
78.S. Motiian, Q. Jones, S. Iranmanesh, and G. Doretto. Few-shot adversarial domain adaptation. In Advances in Neural Information Processing Systems, pages 6673–6683, 2017.
79.S. Motiian, M. Piccirilli, D. A. Adjeroh, and G. Doretto. Unified deep supervised domain adaptation and generalization. In The IEEE International Conference on Computer Vision (ICCV), volume 2, 2017.
80.H. V. Nguyen, H. T. Ho, V. M. Patel, and R. Chellappa. Dash-n: Joint hierarchical domain adaptation and feature learning. IEEE Transactions on Image Processing, 24(12):5479–5491, 2015.
81.S. Pachori, A. Deshpande, and S. Raman. Hashing in the zero shot framework with domain adaptation. Neurocomputing, 2017.
82.S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang. Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks, 22(2):199–210, 2011.
83.S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2010.
84.V. M. Patel, R. Gopalan, R. Li, and R. Chellappa. Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine, 32(3):53–69, 2015.
85.K.-C. Peng, Z. Wu, and J. Ernst. Zero-shot deep domain adaptation. arXiv preprint arXiv:1707.01922, 2017.
86.X. Peng, J. Hoffman, X. Y. Stella, and K. Saenko. Fine-to-coarse knowledge transfer for low-res image classification. In Image Processing (ICIP), 2016 IEEE International Conference on, pages 3683–3687. IEEE, 2016.
87.X. Peng and K. Saenko. Synthetic to real adaptation with deep generative correlation alignment networks. arXiv preprint arXiv:1701.05524, 2017.
88.A. Raj, V. P. Namboodiri, and T. Tuytelaars. Subspace alignment based domain adaptation for rcnn detector. arXiv preprint arXiv:1507.05578, 2015.
89.S.-A. Rebuffi, H. Bilen, and A. Vedaldi. Learning multiple visual domains with residual adapters. arXiv preprint arXiv:1705.08045, 2017.
90.S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee. Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396, 2016.
91.S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99, 2015.
92.M. Rochan and Y. Wang. Weakly supervised localization of novel objects using appearance transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4315–4324, 2015.
93.O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015.
94.S. Rota Bulo and P. Kontschieder. Neural decision forests for semantic image labelling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 81–88, 2014.
95.A. Rozantsev, M. Salzmann, and P. Fua. Beyond sharing weights for deep domain adaptation. arXiv preprint arXiv:1603.06432, 2016.
96.A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick, K. Kavukcuoglu, R. Pascanu, and R. Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016.
97.K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting visual category models to new domains. In European conference on computer vision, pages 213–226. Springer, 2010.
98.K. Saito, Y. Ushiku, and T. Harada. Asymmetric tri-training for unsupervised domain adaptation. arXiv preprint arXiv:1702.08400, 2017.
99.K. Saito, K. Watanabe, Y. Ushiku, and T. Harada. Maximum classifier discrepancy for unsupervised domain adaptation. arXiv preprint arXiv:1712.02560, 2017.
100.S. Sankaranarayanan, Y. Balaji, A. Jain, S. N. Lim, and R. Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. 2017.
101.L. Shao, F. Zhu, and X. Li. Transfer learning for visual categorization: A survey. IEEE transactions on neural networks and learning systems, 26(5):1019–1034, 2015.
102.J. Shen, Y. Qu, W. Zhang, and Y. Yu. Wasserstein distance guided representation learning for domain adaptation. 2017.
103.W. Shimoda and K. Yanai. Distinct class-specific saliency maps for weakly supervised semantic segmentation. In European Conference on Computer Vision, pages 218–234. Springer, 2016.
104.A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and R. Webb. Learning from simulated and unsupervised images through adversarial training. arXiv preprint arXiv:1612.07828, 2016.
105.X. Shu, G.-J. Qi, J. Tang, and J. Wang. Weakly-shared deep transfer networks for heterogeneous-domain knowledge propagation. In Proceedings of the 23rd ACM international conference on Multimedia, pages 35–44. ACM, 2015.
106.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
107.K. Sohn, S. Liu, G. Zhong, X. Yu, M.-H. Yang, and M. Chandraker. Unsupervised domain adaptation for face recognition in unlabeled videos. arXiv preprint arXiv:1708.02191, 2017.
108.B. Sun, J. Feng, and K. Saenko. Return of frustratingly easy domain adaptation. In AAAI, volume 6, page 8, 2016.
109.B. Sun and K. Saenko. Deep coral: Correlation alignment for deep domain adaptation. In Computer Vision–ECCV 2016 Workshops, pages 443–450. Springer, 2016.
110.C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015.
111.Y. Taigman, A. Polyak, and L. Wolf. Unsupervised cross-domain image generation. arXiv preprint arXiv:1611.02200, 2016.
112.Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701–1708, 2014.
113.B. Tan, Y. Song, E. Zhong, and Q. Yang. Transitive transfer learning. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1155–1164. ACM, 2015.
114.B. Tan, Y. Zhang, S. J. Pan, and Q. Yang. Distant domain transfer learning. In AAAI, pages 2604–2610, 2017.
115.Y. Tang, J. Wang, B. Gao, E. Dellandréa, R. Gaizauskas, and L. Chen. Large scale semi-supervised object detection using visual and semantic knowledge transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2119–2128, 2016.
116.J.-C. Tsai and J.-T. Chien. Adversarial domain separation and adaptation. In Machine Learning for Signal Processing (MLSP), 2017 IEEE 27th International Workshop on, pages 1–6. IEEE, 2017.
117.E. Tzeng, C. Devin, J. Hoffman, C. Finn, P. Abbeel, S. Levine, K. Saenko, and T. Darrell. Adapting deep visuomotor representations with weak pairwise constraints. CoRR, vol. abs/1511.07111, 2015.
118.E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and tasks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4068–4076, 2015.
119.E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial discriminative domain adaptation. arXiv preprint arXiv:1702.05464, 2017.
120.E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014.
121.D. Ulyanov, A. Vedaldi, and V. Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. arXiv preprint arXiv:1701.02096, 2017.
122.P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research, 11(Dec):3371–3408, 2010.
123.R. Volpi, P. Morerio, S. Savarese, and V. Murino. Adversarial feature augmentation for unsupervised domain adaptation. arXiv preprint arXiv:1711.08561, 2017.
124.C. Wang and S. Mahadevan. Heterogeneous domain adaptation using manifold alignment. In IJCAI proceedings-international joint conference on artificial intelligence, volume 22, page 1541, 2011.
125.L. Wang, V. A. Sindagi, and V. M. Patel. High-quality facial photosketch synthesis using multi-adversarial networks. arXiv preprint arXiv:1710.10182, 2017.
126.X. Wang, X. Duan, and X. Bai. Deep sketch feature for cross-domain image retrieval. Neurocomputing, 207:387–397, 2016.
127.Y. Xia, D. Huang, and Y. Wang. Detecting smiles of young children via deep transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1673–1681, 2017.
128.T. Xiao, H. Li, W. Ouyang, and X. Wang. Learning deep feature representations with domain guided dropout for person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1249–1258, 2016.
129.M. Xie, N. Jean, M. Burke, D. Lobell, and S. Ermon. Transfer learning from deep features for remote sensing and poverty mapping. 2015.
130.H. Yan, Y. Ding, P. Li, Q. Wang, Y. Xu, and W. Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. arXiv preprint arXiv:1705.00609, 2017.
131.Z. Yi, H. Zhang, P. T. Gong, et al. Dualgan: Unsupervised dual learning for image-to-image translation. arXiv preprint arXiv:1704.02510, 2017.
132.D. Yoo, N. Kim, S. Park, A. S. Paek, and I. S. Kweon. Pixel-level domain transfer. In European Conference on Computer Vision, pages 517–532. Springer, 2016.
133.J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? In Advances in neural information processing systems, pages 3320–3328, 2014.
134.W. Zellinger, T. Grubinger, E. Lughofer, T. Natschläger, and S. Saminger-Platz. Central moment discrepancy (cmd) for domain-invariant representation learning. arXiv preprint arXiv:1702.08811, 2017.
135.H. Zhang, T. Xu, H. Li, S. Zhang, X. Huang, X. Wang, and D. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In IEEE Int. Conf. Comput. Vision (ICCV), pages 5907–5915, 2017.
136.J. Zhang, Z. Ding, W. Li, and P. Ogunbona. Importance weighted adversarial nets for partial domain adaptation. arXiv preprint arXiv:1803.09210, 2018.
137.J. Zhang, W. Li, and P. Ogunbona. Transfer learning for cross-dataset recognition: A survey. 2017.
138.L. Zhang, Z. He, and Y. Liu. Deep object recognition across domains based on adaptive extreme learning machine. Neurocomputing, 239:194–203, 2017.
139.X. Zhang, F. X. Yu, S.-F. Chang, and S. Wang. Deep transfer network: Unsupervised domain adaptation. arXiv preprint arXiv:1503.00591, 2015.
140.Y. Zhang, P. David, and B. Gong. Curriculum domain adaptation for semantic segmentation of urban scenes. In The IEEE International Conference on Computer Vision (ICCV), volume 2, page 6, 2017.
141.W. Zhao, W. Xu, M. Yang, J. Ye, Z. Zhao, Y. Feng, and Y. Qiao. Dual learning for cross-domain image captioning. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 29–38. ACM, 2017.
142.J. T. Zhou, I. W. Tsang, S. J. Pan, and M. Tan. Heterogeneous domain adaptation for multiple classes. In Artificial Intelligence and Statistics, pages 1095–1103, 2014.
143.J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
144.F. Zhuang, X. Cheng, P. Luo, S. J. Pan, and Q. He. Supervised representation learning: Transfer learning with deep autoencoders. In IJCAI, pages 4119–4125, 2015.

Citation

MLA

Wang, M., and W. Deng. “Deep Visual Domain Adaptation: A Survey”. Neurocomputing, vol. 312, 2018, pp. 135–53, https://doi.org/10.1016/J.NEUCOM.2018.05.083.

APA

Wang, M., & Deng, W. (2018). Deep visual domain adaptation: A survey. Neurocomputing, 312, 135–153. https://doi.org/10.1016/J.NEUCOM.2018.05.083

Chicago

Wang, M., and W. Deng. 2018. “Deep Visual Domain Adaptation: A Survey”. Neurocomputing 312: 135–53. https://doi.org/10.1016/J.NEUCOM.2018.05.083.

Harvard

Wang, M. and Deng, W. (2018) “Deep visual domain adaptation: A survey”, Neurocomputing, 312, pp. 135–153. Available at: https://doi.org/10.1016/J.NEUCOM.2018.05.083.

Vancouver

1. Wang M, Deng W (2018) Deep visual domain adaptation: A survey. Neurocomputing 312:135–153

BibTeX

@article{Wang_2018, title={Deep visual domain adaptation: A survey}, volume={312}, ISSN={0925-2312}, url={http://dx.doi.org/10.1016/J.NEUCOM.2018.05.083}, DOI={10.1016/j.neucom.2018.05.083}, journal={Neurocomputing}, publisher={Elsevier BV}, author={Wang, Mei and Deng, Weihong}, year={2018}, month=Oct, pages={135–153} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Deep Visual Domain Adaptation: A Survey

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Taxonomy and Formal Problem Settings of Visual Domain Adaptation

Knowl 2 — Taxonomy of One-Step Deep Domain Adaptation Approaches

Knowl 3 — Statistical Distribution Alignment in Deep Domain Adaptation

Knowl 4 — Class, Architectural, and Geometric Alignment Criteria in Deep DA

Knowl 5 — Adversarial-Based Deep Domain Adaptation

Knowl 6 — Reconstruction-Based Deep Domain Adaptation

Knowl 7 — Heterogeneous Deep Domain Adaptation Mechanisms

Knowl 8 — Mechanisms for Multi-Step (Transitive) Deep Domain Adaptation

Knowl 9 — Empirical Classification Accuracy of Deep DA vs. Non-Adaptation Baselines

Knowl 10 — Current Limitations and Open Frontiers in Deep Domain Adaptation

References

Citation

BibTeX

Access the Paper