C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation

Nazmul KarimNiluthpol Chowdhury MithunAbhinav RajvanshiHan-Pang ChiuSupun SamarasekeraNazanin Rahnavard

article2023CVPR113 citations

Proposes a curriculum-driven self-training framework for source-free domain adaptation that filters out noisy pseudo-labels to prevent early-stage memorization without needing memory banks or expensive feature clustering.

Listen

Deep neural network models often experience steep performance drops when deployed on data that differs from their original training environment. While standard adaptation techniques require ongoing access to original source datasets, real-world constraints such as data privacy regulations, proprietary restrictions, and limited hardware on edge devices frequently make source data unavailable. Existing source-free methods attempt self-training using model-generated labels, but they suffer from early memorization of incorrect predictions and rely heavily on memory-intensive storage banks to refine labels, making deployment on resource-constrained platforms difficult.

The article introduces and evaluates C-SFDA, a curriculum learning framework designed to achieve accurate, memory-efficient source-free domain adaptation without needing source data or complex memory banks. The approach structures the learning process into a curriculum that prioritizes high-confidence, low-uncertainty target samples first, gradually propagating refined knowledge to harder samples while using unsupervised representation learning to prevent memorizing false predictions. The framework was evaluated across multiple standard image classification and semantic segmentation benchmarks in both offline and online deployment settings.

The experimental findings show that the proposed framework consistently surpasses existing state-of-the-art source-free methods across visual recognition tasks. In image classification, it achieved average accuracy gains of 0.4% on Office-31, 0.6% on Office-Home, 1.0% on VisDA (reaching 87.8%), and 1.2% on DomainNet (reaching 69.0%). In semantic segmentation, it achieved leading accuracy on synthetic-to-real benchmarks, such as 48.3% mean intersection-over-union on Cityscapes, while matching or exceeding methods that require continuous access to original source data. Furthermore, in real-time online adaptation settings where models learn in a single pass, the framework outperformed prior approaches across both classification and segmentation tasks, including a 3.5% accuracy gain on VisDA and a 4.0% segmentation improvement on SYNTHIA-to-Cityscapes.

These results demonstrate that complex, memory-heavy label storage mechanisms are unnecessary for effective domain transfer. Organizations can reliably adapt artificial intelligence systems to new operational domains with lower computational costs, lower memory overhead, and strict adherence to data privacy requirements. The strategy of filtering out noisy early predictions enables stable self-training even under severe domain shifts.

Organizations deploying computer vision models in resource-limited or privacy-sensitive environments should consider adopting selective, curriculum-based self-training frameworks. However, the authors note that initial label reliability can degrade when domain shifts are exceptionally severe, which may lead to overly restrictive sample selection. In such extreme cases, decision-makers should consider pairing this approach with robust self-supervised pre-training or strongly augmented source-model preparation to ensure stable adaptation.

SRI International University of Central Florida curricula source-free domain adaptation curriculum learning self-training framework

arXiv: 2303.17132

Paper: Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation, Jian Liang et al. (2020). This seminal work establishes the Source Hypothesis Transfer (SHOT) framework for source-free domain adaptation via information maximization and pseudo-labeling, which C-SFDA directly builds upon and enhances.
Paper: FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling, Bowen Zhang et al. (2021). This paper introduces Curriculum Pseudo Labeling to adjust sample selection thresholds based on model learning status, providing the foundational curriculum self-training strategy adapted by C-SFDA.
Paper: Attracting and Dispersing: A Simple Approach for Source-free Domain Adaptation, Shiqi Yang et al. (2022). It develops an efficient clustering-based source-free domain adaptation method that serves as a direct baseline and context for C-SFDA's memory-efficient curriculum framework.
Paper: Contrastive Test-Time Adaptation, Dian Chen et al. (2022). It proposes test-time adaptation with contrastive learning and memory-queue pseudo-label refinement, offering the key architectural backdrop for C-SFDA's memory-bank-free design.
Paper: Confidence Score for Source-Free Unsupervised Domain Adaptation, Jonghyun Lee et al. (2022). It formulates sample-wise confidence scoring in source-free domain adaptation to combat noisy pseudo-labels, addressing the central problem C-SFDA tackles with curriculum learning.
Paper: Balancing Discriminability and Transferability for Source-Free Domain Adaptation, Jogendra Nath Kundu et al. (2022). It examines balancing discriminability and transferability under strict source-free settings, providing essential theoretical and empirical context for C-SFDA.
Paper: Unsupervised Domain Adaptation for Semantic Segmentation via Class-Balanced Self-training, Yang Zou et al. (2018). This paper establishes iterative self-training and confidence thresholding for domain adaptation across visual recognition and semantic segmentation tasks.
Paper: Continual Test-Time Domain Adaptation, Qin Wang et al. (2022). It explores online and continual test-time domain adaptation without source data, defining the streaming adaptation scenarios evaluated in C-SFDA.

Paper: Robust Mean Teacher for Continual and Gradual Test-Time Adaptation, Mario Döbler et al. (2023). This paper extends source-free test-time adaptation to continual and gradual environment shifts using a mean-teacher architecture with contrastive anchors.
Paper: Feature Alignment and Uniformity for Test Time Adaptation, Shuai Wang et al. (2023). It furthers online test-time adaptation by combining dual pseudo-label filtering with feature uniformity and spatial clustering objectives.
Paper: Instance Relation Graph Guided Source-Free Domain Adaptive Object Detection, Vibashan VS et al. (2023). It extends source-free adaptation methodologies to complex structured object detection using instance relation graphs and proposal-level contrastive learning.
Paper: Improved Test-Time Adaptation for Domain Generalization, Liang Chen et al. (2023). It develops test-time adaptation strategies with lightweight adaptive parameters for out-of-distribution domain generalization.
Paper: Dual Memory Networks: A Versatile Adaptation Approach for Vision-Language Models, Yabin Zhang et al. (2024). It applies online dynamic test-time adaptation principles to large vision-language foundation models via cross-attention dual memory networks.

Cover for C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) approaches focus on adapting models trained on a labeled source domain to an unlabeled target domain. In contrast to UDA, source-free domain adaptation (SFDA) is a more practical setup as access to source data is no longer required during adaptation. Recent state-of-the-art (SOTA) methods on SFDA mostly focus on pseudo-label refinement based self-training which generally suffers from two issues: i) inevitable occurrence of noisy pseudo-labels that could lead to early training time memorization, ii) refinement process requires maintaining a memory bank which creates a significant burden in resource constraint scenarios. To address these concerns, we propose C-SFDA, a curriculum learning aided self-training framework for SFDA that adapts efficiently and reliably to changes across domains based on selective pseudo-labeling. Specifically, we employ a curriculum learning scheme to promote learning from a restricted amount of pseudo labels selected based on their reliabilities. This simple yet effective step successfully prevents label noise propagation during different stages of adaptation and eliminates the need for costly memory-bank based label refinement. Our extensive experimental evaluations on both image recognition and semantic segmentation tasks confirm the effectiveness of our method. C-SFDA is also applicable to online test-time domain adaptation and outperforms previous SOTA methods in this task.

Topics

curricula

Keywords

source-free domain adaptation curriculum learning self-training framework pseudo-label refinement label noise memorization label propagation loss unsupervised domain adaptation test-time domain adaptation domain shift

1. Introduction
2. Related Work
3. Proposed Method
3.1. Curriculum SFDA
3.1.1. Selective Pseudo-Labelling
3.1.2. Loss Functions
3.1.3. Curriculum Learning
3.1.4. Semantic Segmentation
4. Experiments
4.1. Experimental Settings
4.2. Experimental Results
4.2.1. Ablation Study
5. Conclusion
References

Knowls

Knowl 1 — Teacher-Student Self-Training Framework with Augmentation and Weight Averaging in C-SFDA
model/method
In source-free domain adaptation (SFDA), target domain adaptation is performed without access to labeled source domain data $\mathcal{D}_s = \{(x_s^i, y_s^i)\}_{i=1}^{N_s}$ , utilizing only a source pre-trained model $f_{\theta_s}$ and an unlabeled target domain dataset $\mathcal{D}_t = \{x_t^i\}_{i=1}^{N_t}$ across $K$ shared semantic categories.

C-SFDA establishes a self-training framework with a student network $f_{\theta_t}$ and a teacher network $f_{\hat{\theta}_t}$ . Both networks are initialized with source weights: $\theta_t = \hat{\theta}_t = \theta_s$ . To stabilize predictions and mitigate label noise without requiring memory banks or feature clustering, target pseudo-labels are generated by averaging the teacher model's predictions over $L$ stochastic augmentations of each target input $x_t$ :

$\hat{y}_t = \arg\max \frac{1}{L} \sum_{l=1}^{L} \hat{h}_t^l = \arg\max \frac{1}{L} \sum_{l=1}^{L} f_{\hat{\theta}_t}(\hat{x}_t^l)$

where $\hat{x}_t^l$ denotes the $l$ -th augmented version of $x_t$ , $\hat{h}_t^l = f_{\hat{\theta}_t}(\hat{x}_t^l) \in \mathbb{R}^K$ is the predicted probability distribution, and $L = 12$ .

At each training iteration step $j \to j+1$ , the teacher model parameters $\hat{\theta}_t$ are updated via an Exponential Moving Average (EMA) of the student weights $\theta_t$ :

$\hat{\theta}_t^{j+1} = \gamma \hat{\theta}_t^j + (1 - \gamma) \theta_t^{j+1}$

where $\gamma \in [0, 1]$ is a momentum smoothing factor (set to $\gamma = 0.98$ for image classification and $\gamma = 0.995$ for semantic segmentation).
Knowl 2 — Adaptive Confidence and Uncertainty Thresholding for Selective Pseudo-Labeling
model/method
To prevent early training-time memorization (ETM) of noisy pseudo-labels, C-SFDA computes sample-level reliability using prediction confidence and aleatoric prediction uncertainty.

For each target sample $x_t^i$ , prediction confidence is defined as the maximum class probability of the augmentation-averaged prediction $\hat{h}_t^i = \frac{1}{L} \sum_{l=1}^L \hat{h}_t^{l,i}$ :

$\text{conf}(\hat{h}_t^i) = \max_{k \in \{1,\dots,K\}} \hat{h}_{t,k}^i$

Aleatoric uncertainty $g_u^i$ is measured as the standard deviation of prediction confidences across the $L$ augmentations:

$g_u^i = \text{std}\left(\{\text{conf}(f_{\hat{\theta}_t}(\hat{x}_t^{l,i}))\}_{l=1}^L\right)$

For a target mini-batch of size $B$ , adaptive selection thresholds $\tau_c$ (confidence) and $\tau_u$ (uncertainty) are computed dynamically as the batch means:

$\tau_c = \frac{1}{B} \sum_{i=1}^B \text{conf}(\hat{h}_t^i), \qquad \tau_u = \frac{1}{B} \sum_{i=1}^B g_u^i$

A binary reliability indicator $r^i \in \{0, 1\}$ is assigned to sample $x_t^i$ via:

$r^i = \begin{cases} 1, & \text{if } \text{conf}(\hat{h}_t^i) \ge \tau_c \text{ and } g_u^i \le \tau_u \\ 0, & \text{otherwise} \end{cases}$

The mini-batch is thereby split into a reliable subset $\mathbb{D}_R = \{(x_t^i, \hat{y}_t^i) : r^i = 1\}$ and an unreliable subset $\mathbb{D}_U = \{(x_t^i, \hat{y}_t^i) : r^i = 0\}$ .
Knowl 3 — Composite Classification Adaptation Objective in C-SFDA
equation
The overall adaptation objective for image classification in C-SFDA combines supervised learning on reliable pseudo-labels, label propagation on unreliable samples, and unsupervised contrastive representation learning.
1. Class-Balanced Cross-Entropy Loss on $\mathbb{D}_R$ : $\mathcal{L}_{ce}^R = - \frac{1}{|\mathbb{D}_R|} \sum_{i=1}^{|\mathbb{D}_R|} \lambda_{\hat{y}_t^i} \, \hat{y}_{tc}^i \cdot \log f_{\theta_t}(x_t^i)$ where $\hat{y}_{tc}^i$ is the one-hot encoding of pseudo-label $\hat{y}_t^i$ , and $\lambda_k$ is an inverse class-frequency weighting factor balancing the distribution in $\mathbb{D}_R$ .
2. Label Propagation Loss on $\mathbb{D}_U$ : $\mathcal{L}_{\mathcal{P}} = \frac{1}{2|\mathbb{D}_U|} \sum_{i=1}^{|\mathbb{D}_U|} \| f_{\theta_t}(x_t^i) - \hat{y}_{tc}^i \|_2^2$ which acts as a transductive regularizer diffusing learned representations to less confident samples.
3. Unsupervised Contrastive Loss on target features: Given two augmented views $x_t^{\text{aug},1}$ and $x_t^{\text{aug},2}$ projected via feature extractor $G$ and projection head $H$ to $q_i = H(G(x_t^{\text{aug},1}))$ and $q_j = H(G(x_t^{\text{aug},2}))$ : $\ell_{i,j} = -\log \frac{\exp(\text{sim}(q_i, q_j)/\kappa)}{\sum_{b=1}^{2B} \mathbf{1}_{b \neq i} \exp(\text{sim}(q_i, q_b)/\kappa)}$ $\mathcal{L}_{\mathcal{C}} = \frac{1}{2B} \sum_{b=1}^{2B} [\ell_{2b-1, 2b} + \ell_{2b, 2b-1}]$ where $\text{sim}(q_i, q_j) = \frac{q_i^\top q_j}{\|q_i\| \|q_j\|}$ is cosine similarity, $\kappa$ is a temperature hyperparameter, and $\mathbf{1}_{b \neq i}$ is an indicator function.
The composite objective at iteration $j$ is: $\mathcal{L}_{tot} = \mu_r^j \mathcal{L}_{ce}^R + (1 - \mu_r^j) \mathcal{L}_{\mathcal{P}} + \mu_c^j \mathcal{L}_{\mathcal{C}}$ where $\mu_r^j$ and $\mu_c^j$ are curriculum-controlled loss weighting coefficients.
Knowl 4 — Difficulty-Aware Curriculum Learning Pacing Schedule
equation
To dynamically modulate learning between reliable self-training, label propagation, and unsupervised contrastive representation learning, C-SFDA defines a batch difficulty score $d^j$ at iteration $j$ as the ratio of average uncertainty to average confidence:

$d^j = \frac{\tau_u^j}{\tau_c^j}$

where $\tau_u^j$ and $\tau_c^j$ are the mean batch uncertainty and confidence thresholds.

The reliable loss coefficient $\mu_r^j$ is updated across iterations as:

$\mu_r^j = \mu_r^{j-1}\left(1 - \alpha \exp\left(-\frac{1}{d^j}\right)\right)$

with initial value $\mu_r^0 = 1.0$ and step decay rate $\alpha = 0.005$ . When a batch is hard to learn ( $d^j$ is large), $\mu_r$ decays minimally, constraining learning to verified reliable samples. As domain adaptation progresses and $d^j$ drops, $\mu_r$ decreases, gradually increasing the weight of label propagation $(1 - \mu_r^j)$ on $\mathbb{D}_U$ .

The contrastive loss coefficient $\mu_c^j$ decays exponentially:

$\mu_c^j = \mu_c^{j-1} e^{-\beta}$

with initial value $\mu_c^0 = 0.5$ and decay rate $\beta = 10^{-4}$ , providing strong label-independent feature regularization early in training to prevent premature label noise memorization.
Knowl 5 — Selective Pseudo-Labeling and Loss for Semantic Segmentation
model/method
For semantic segmentation on target domain image $x \in \mathbb{R}^{H \times W}$ , the network predicts pixel-wise class probability maps $p \in \mathbb{R}^{H \times W \times K}$ , yielding pseudo-labels $\hat{y}_{ij} = \arg\max_k p_{ij,k}$ .

To account for extreme class imbalances across spatial dimensions, C-SFDA computes per-category selection thresholds. For each category $k \in \{1, \dots, K\}$ , the confidence threshold $\tau_c^k$ and uncertainty threshold $\tau_u^k$ (computed using ColorJitter and Gaussian noise augmentations) are set to the $P$ -th percentile of confidence and uncertainty values accumulated over the batch for class $k$ , with $P = 55$ .

A pixel at location $(i, j)$ is labeled reliable ( $r_{ij}=1$ ) if its confidence is at least $\tau_c^{\hat{y}_{ij}}$ and uncertainty is at most $\tau_u^{\hat{y}_{ij}}$ .

During adaptation, all network weights are frozen except for the Batch Normalization (BN) layers, which are optimized via:

$\mathcal{L}_{tot} = \mathcal{L}_{ce}^R + \mu_e \mathcal{L}_E$

where $\mathcal{L}_{ce}^R$ is the cross-entropy loss over reliable pixels, $\mu_e$ is an entropy weight initialized to $\mu_e^0 = 10^{-3}$ and updated via $\mu_e^j = \mu_e^{j-1} e^{-\beta}$ , and $\mathcal{L}_E$ is the mean pixel-wise entropy:

$\mathcal{L}_E = - \frac{1}{HW} \sum_{i=1}^H \sum_{j=1}^W \sum_{k=1}^K p_{ij,k} \log(p_{ij,k})$

Knowl 6 — Classification Performance on Office-31 and Office-Home SFDA Benchmarks

data/table

Evaluations of C-SFDA on Office-31 (3 domains, 31 classes, ResNet-50) and Office-Home (4 domains, 65 classes, ResNet-50) demonstrate superior accuracy over prior source-free domain adaptation (SFDA) and unsupervised domain adaptation (UDA) methods.

Office-31 Classification Accuracy (%)
Method	SF	A $\to$ D	A $\to$ W	D $\to$ A	D $\to$ W	W $\to$ A	W $\to$ D	Avg.
GSDA	$\times$	94.8	95.7	73.5	99.1	74.9	100.0	89.7
CAN	$\times$	95.0	94.5	78.0	99.1	77.0	99.8	90.6
SRDC	$\times$	95.8	95.7	76.7	99.2	77.1	100.0	90.8
SHOT	✓	94.0	90.1	74.7	98.4	74.3	99.9	88.6
3C-GAN	✓	92.7	93.7	75.3	98.5	77.8	99.8	89.6
$A^2$ Net	✓	94.5	94.0	76.7	99.2	76.1	100.0	90.1
SFDA-DE	✓	96.0	94.2	76.6	98.5	75.5	99.8	90.1
C-SFDA (Ours)	✓	96.2	93.9	77.3	98.8	77.9	99.7	90.5
Office-Home Performance Summary (12 Shifts)
FixBi (UDA): 72.7%		SHOT (SFDA): 71.8%			HCL (SFDA): 72.6%
$A^2$ Net (SFDA): 72.8%		SFDA-DE (SFDA): 72.9%			C-SFDA (Ours): 73.5%

C-SFDA achieves an average accuracy of 90.5% on Office-31 (+0.4% over previous SOTA SFDA-DE) and 73.5% on Office-Home (+0.6% over SFDA-DE, achieving SOTA accuracy on 8 of the 12 domain transfer tasks).

Knowl 7 — Classification Performance on VisDA and DomainNet in Offline and Online Adaptation

data/table

Performance comparison on the VisDA-2017 synthetic-to-real benchmark (ResNet-101) and DomainNet benchmark (ResNet-50) under both standard offline SFDA and single-epoch online test-time domain adaptation.

Method	Source-Free	Online	VisDA Avg. Acc. (%)	DomainNet Avg. Acc. (%)
Source only	–	–	43.8	55.6
MCC	$\times$	–	78.8	48.9
SHOT	✓	–	82.9	67.1
$A^2$ Net	✓	–	84.3	–
SFDA-DE	✓	–	86.5	–
AdaCon	✓	–	86.8	67.8
C-SFDA (Ours)	✓	–	87.8	69.0
TENT	✓	✓	–	57.7
AdaCon (Online)	✓	✓	78.7	62.6
C-SFDA (Online)	✓	✓	82.1	63.1

In standard offline SFDA, C-SFDA attains 87.8% on VisDA (+1.0% over AdaCon, with accuracy on the rare 'truck' class rising from 49.7% to 63.5%) and 69.0% on DomainNet (+1.2% over AdaCon). In online test-time adaptation, C-SFDA achieves 82.1% on VisDA (+3.4% over AdaCon Online) and 63.1% on DomainNet (+0.5% over AdaCon Online).

Knowl 8 — Semantic Segmentation Performance on GTA5, SYNTHIA, and Dark-Zurich

data/table

Evaluation of C-SFDA on semantic segmentation benchmarks using DeepLabV2 with a ResNet-101 backbone on synthetic-to-real (GTA5 $\to$ Cityscapes, SYNTHIA $\to$ Cityscapes) and day-to-night (Cityscapes $\to$ Dark-Zurich) adaptation.

Method	SF	Online	GTA5 $\to$ Cityscapes (19-way mIoU %)	SYNTHIA $\to$ Cityscapes (16-way / 13-way* mIoU %)
Source Only	–	–	36.4	31.3 / 36.2
CrCDA (UDA)	$\times$	–	48.6	42.9 / 50.0
UR	✓	–	45.1	39.6 / 45.0
SFDA	✓	–	45.8	42.4 / 48.7
HCL	✓	–	48.1	43.5 / 50.2
C-SFDA (Ours)	✓	–	48.3	44.6 / 51.3
AUGCO (Online)	✓	✓	45.9	39.2 / 45.5
C-SFDA (Online)	✓	✓	46.3	43.0 / 49.5
Cityscapes $\to$ Dark-Zurich (19-way mIoU %)
TTBN: 28.8 TENT: 28.0 AUGCO: 32.4 C-SFDA (Ours): 33.2

C-SFDA achieves state-of-the-art SFDA results across all benchmarks without requiring a memory queue, outperforming HCL by +0.2% on GTA5 $\to$ Cityscapes (48.3%) and by +1.1% on SYNTHIA $\to$ Cityscapes (44.6% 16-way / 51.3% 13-way). In online adaptation (1 epoch), C-SFDA outperforms AUGCO by +0.4% on GTA5 $\to$ Cityscapes, +3.8% on SYNTHIA $\to$ Cityscapes, and +0.8% on Cityscapes $\to$ Dark-Zurich (33.2%).

Knowl 9 — Ablation of Selection Criteria and Loss Components in C-SFDA

empirical result

Ablation experiments across Office-31, Office-Home, VisDA, and DomainNet isolate the contribution of individual components in the C-SFDA framework:

Selection Strategy			Label Bal.	Loss		Accuracy (%)
Conf.	Unc.	DoC	$\lambda_k$	$\mathcal{L}_{\mathcal{P}}$	$\mathcal{L}_{\mathcal{C}}$	Office-31	Office-Home	VisDA	DomainNet
Self-training ( $\mathcal{L}_{ce}$ ) with all pseudo-labels						81.1	62.3	57.2	52.6
✓	–	–	✓	✓	✓	87.6	69.2	85.2	65.5
✓	✓	–	✓	✓	✓	89.9	71.8	87.4	68.7
✓	✓	✓	–	✓	✓	90.1	73.3	86.5	68.3
✓	✓	✓	✓	–	–	88.7	71.6	85.9	67.3
✓	✓	✓	✓	✓	–	88.9	72.3	86.4	67.9
✓	✓	✓	✓	✓	✓	90.5	73.5	87.8	69.0

Key observations:

Unfiltered self-training on all pseudo-labels leads to early memorization of label noise, collapsing accuracy on VisDA to 57.2% and DomainNet to 52.6%.
Incorporating uncertainty filtering with confidence thresholding improves VisDA accuracy from 85.2% to 87.4% and DomainNet from 65.5% to 68.7%.
Combining label propagation $\mathcal{L}_{\mathcal{P}}$ with unsupervised contrastive learning $\mathcal{L}_{\mathcal{C}}$ is essential; without contrastive learning, performance drops from 87.8% to 86.4% on VisDA.
Applying the dynamic curriculum schedule yields +0.7% on VisDA (87.8% vs. 87.1% without curriculum) and +0.4% on DomainNet (69.0% vs. 68.6%).

Knowl 10 — Failure of Traditional Label Noise Learning Techniques in SFDA
empirical result
Traditional Label Noise Learning (LNL) methods designed for supervised learning with noisy labels (e.g., Generalized Cross Entropy [GCE], Progressive Curriculum Learning [PCL], and Early-Learning Regularization [ELR]) fail when applied to Source-Free Domain Adaptation.

Empirical comparisons across datasets show:
- Standard cross-entropy with all pseudo-labels yields ~62.3% on Office-Home, ~57.2% on VisDA, and ~52.6% on DomainNet.
- LNL methods (GCE, PCL, ELR) achieve only minor improvements, remaining under 70% on VisDA and under 60% on DomainNet.
- In contrast, C-SFDA reaches 73.5% on Office-Home, 87.8% on VisDA, and 69.0% on DomainNet.
This performance deficit occurs because conventional LNL algorithms assume bounded label noise where noise rates or types are constrained. In SFDA, domain shift causes unbounded label noise where noise rates can be exceptionally high and class-dependent without known prior distributions, rendering standard noise regularization ineffective.
Knowl 11 — Sensitivity to Initial Pseudo-Label Quality Under Severe Domain Shifts
limitation
The selective pseudo-labeling in C-SFDA depends on prediction confidence and augmentation-derived uncertainty from the source-initialized model. In scenarios where the domain shift between source and target domains is exceptionally large, initial pseudo-labels can be severely inaccurate and unreliable. Under these extreme conditions, the adaptive batch-average thresholds can become overly restrictive, selecting an insufficient number of reliable samples for supervised self-training, or may select spuriously overconfident incorrect predictions. Addressing such extreme domain shifts requires complementary techniques, such as noise-robust self-supervised pre-training or aggressive source-domain data augmentation.

Coverage note — No substantial contributed material was omitted; the knowls cover the full C-SFDA framework, selective pseudo-labeling criteria, composite loss equations, curriculum schedule, segmentation formulation, experimental results across all benchmarks, ablations, comparison with LNL, and limitations.

References

1.Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233–242. PMLR, 2017.
2.Yoshua Bengio, Jer´ ome Louradour, Ronan Collobert, and Ja- ˆ son Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009.
3.Chao Chen, Zhihang Fu, Zhihong Chen, Sheng Jin, Zhaowei Cheng, Xinyu Jin, and Xian-Sheng Hua. Homm: Higher-order moment matching for unsupervised domain adaptation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3422–3429, 2020.
4.Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295–305, 2022.
5.Liqun Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. Graph optimal transport for cross-domain alignment. In International Conference on Machine Learning, pages 1542–1553. PMLR, 2020.
6.Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017.
7.Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020.
8.Jaehoon Choi, Minki Jeong, Taekyung Kim, and Changick Kim. Pseudo-labeling curriculum for unsupervised domain adaptation. arXiv preprint arXiv:1908.00262, 2019.
9.Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016.
10.Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. Domain adaptation in computer vision applications, pages 1–35, 2017.
11.Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113–123, 2019.
12.Bharath Bhushan Damodaran, Benjamin Kellenberger, Remi ´ Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447–463, 2018.
13.Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009.
14.Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7212–7222, 2022.
15.Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. Advances in Neural Information Processing Systems, 34:2848–2860, 2021.
16.Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.
17.Hao Feng, Minghao Chen, Jinming Hu, Dong Shen, Haifeng Liu, and Deng Cai. Complementary pseudo labels for unsupervised domain adaptation on person re-identification. IEEE Transactions on Image Processing, 30:2898–2907, 2021.
18.Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation. In International Conference on Learning Representations, number 6, 2018.
19.Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
20.Devin Guillory, Vaishaal Shankar, Sayna Ebrahimi, Trevor Darrell, and Ludwig Schmidt. Predicting with confidence on unseen distributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1134–1144, 2021.
21.Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020.
22.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
23.Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989–1998. Pmlr, 2018.
24.Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Unsupervised domain adaptation with hierarchical gradient synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4043–4052, 2020.
25.Ran He Hu and Jiashi Feng. A balanced and uncertainty-aware approach for partial domain adaptation.
26.Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. Advances in Neural Information Processing Systems, 34:3635–3649, 2021.
27.Jiaxing Huang, Shijian Lu, Dayan Guan, and Xiaobing Zhang. Contextual-relation consistent domain adaptation for semantic segmentation. In European conference on computer vision, pages 705–722. Springer, 2020.
28.Eyke Hullermeier and Willem Waegeman. Aleatoric and ¨ epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine Learning, 110(3):457–506, 2021.
29.Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448–456. PMLR, 2015.
30.Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In European Conference on Computer Vision, pages 464–480. Springer, 2020.
31.Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4893–4902, 2019.
32.Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661–18673, 2020.
33.Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017.
34.Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544–4553, 2020.
35.Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12376–12385, 2020.
36.Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Diverse image-to-image translation via disentangled representations. In Proceedings of the European conference on computer vision (ECCV), pages 35–51, 2018.
37.Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641–9650, 2020.
38.Ruihuang Li, Shuai Li, Chenhang He, Yabin Zhang, Xu Jia, and Lei Zhang. Class-balanced pixel-level self-labeling for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11593–11603, 2022.
39.Shuang Li, Chi Liu, Qiuxia Lin, Binhui Xie, Zhengming Ding, Gao Huang, and Jian Tang. Domain conditioned adaptation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11386–11393, 2020.
40.Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11516–11525, 2021.
41.Justin Liang, Namdar Homayounfar, Wei-Chiu Ma, Yuwen Xiong, Rui Hu, and Raquel Urtasun. Polytransform: Deep polygon transformer for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9131–9140, 2020.
42.Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028–6039. PMLR, 2020.
43.Hong Liu, Jianmin Wang, and Mingsheng Long. Cycle self-training for domain adaptation. Advances in Neural Information Processing Systems, 34:22968–22981, 2021.
44.Sheng Liu, Kangning Liu, Weicheng Zhu, Yiqiu Shen, and Carlos Fernandez-Granda. Adaptive early-learning correction for segmentation from noisy annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2606–2616, 2022.
45.Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331–20342, 2020.
46.Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1215–1224, 2021.
47.Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Advances in neural information processing systems, 31, 2018.
48.Zhihe Lu, Yongxin Yang, Xiatian Zhu, Cong Liu, Yi-Zhe Song, and Tao Xiang. Stochastic classifiers for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9111–9120, 2020.
49.Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In European conference on computer vision, pages 415–430. Springer, 2020.
50.Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019.
51.Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4500–4509, 2018.
52.Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. Fixbi: Bridging domain spaces for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1094–1103, June 2021.
53.Zachary Nado, Shreyas Padhy, D Sculley, Alexander D’Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.
54.Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406–1415, 2019.
55.Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.
56.Viraj Uday Prabhu, Shivam Khare, Deeksha Kartik, and Judy Hoffman. Augmentation consistency-guided self-training for source-free domain adaptive semantic segmentation. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications.
57.Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6790–6800, 2021.
58.Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102–118. Springer, 2016.
59.German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234–3243, 2016.
60.Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In European conference on computer vision, pages 213–226. Springer, 2010.
61.Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8050–8058, 2019.
62.Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7374–7383, 2019.
63.Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. Advances in neural information processing systems, 29, 2016.
64.Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.
65.Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1–48, 2019.
66.Prabhu Teja Sivaprasad and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In 2021 Ieee/Cvf Conference On Computer Vision And Pattern Recognition, Cvpr 2021, number CONF, pages 9608–9618. IEEE, 2021.
67.Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443–450. Springer, 2016.
68.Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725–8735, 2020.
69.Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. Vdm-da: Virtual domain modeling for source data-free domain adaptation. arXiv preprint arXiv:2103.14357, 2021.
70.Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167–7176, 2017.
71.Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018–5027, 2017.
72.Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Perez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517–2526, 2019.
73.Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021.
74.Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135–153, 2018.
75.Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudo-labeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6243–6250, 2020.
76.Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211, 2022.
77.Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9092–9101, 2021.
78.Haifeng Xia and Zhengming Ding. Structure preserving generative cross-domain learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4364–4373, 2020.
79.Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9010–9019, 2021.
80.Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687–10698, 2020.
81.Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4394–4403, 2020.
82.Yixing Xu, Kai Han, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Learning frequency domain approximation for binary neural networks. Advances in Neural Information Processing Systems, 34, 2021.
83.Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. Advances in Neural Information Processing Systems, 34:29393–29405, 2021.
84.Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 1(2):5, 2020.
85.Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8978–8987, 2021.
86.Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10754–10762, 2021.
87.Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12414–12424, 2021.
88.Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756, 2021.
89.Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018.
90.Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and ¨ global consistency. Advances in neural information processing systems, 16, 2003.
91.Tianyi Zhou and Jeff Bilmes. Minimax curriculum learning: Machine teaching with desirable difficulties and scheduled diversity. In International Conference on Learning Representations, 2018.
92.Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289–305, 2018.
93.Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982–5991, 2019.

Citation

MLA

Karim, N., et al. “C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation”. arXiv, 2023, http://arxiv.org/abs/2303.17132v1.

APA

Karim, N., Mithun, N. C., Rajvanshi, A., Chiu, H.-. pang ., Samarasekera, S., & Rahnavard, N. (2023). C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation. arXiv. http://arxiv.org/abs/2303.17132v1

Chicago

Karim, N., N. C. Mithun, A. Rajvanshi, H.-. pang . Chiu, S. Samarasekera, and N. Rahnavard. 2023. “C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation”. arXiv. http://arxiv.org/abs/2303.17132v1.

Harvard

Karim, N. et al. (2023) “C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2303.17132v1.

Vancouver

1. Karim N, Mithun NC, Rajvanshi A, Chiu H-pang, Samarasekera S, Rahnavard N (2023) C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation. arXiv

BibTeX

@article{karim2023sfda,
  title = {C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation},
  author = {Karim, Nazmul and Mithun, Niluthpol Chowdhury and Rajvanshi, Abhinav and Chiu, Han-pang and Samarasekera, Supun and Rahnavard, Nazanin},
  year = {2023},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2303.17132v1},
  eprint = {2303.17132}
}

Metadata:arXiv

Access the Paper

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

Open PDF

License: IEEE

C-SFDA: A Curriculum Learning Aided Self-Training Framework for Efficient Source Free Domain Adaptation

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Teacher-Student Self-Training Framework with Augmentation and Weight Averaging in C-SFDA

Knowl 2 — Adaptive Confidence and Uncertainty Thresholding for Selective Pseudo-Labeling

Knowl 3 — Composite Classification Adaptation Objective in C-SFDA

Knowl 4 — Difficulty-Aware Curriculum Learning Pacing Schedule

Knowl 5 — Selective Pseudo-Labeling and Loss for Semantic Segmentation

Knowl 6 — Classification Performance on Office-31 and Office-Home SFDA Benchmarks

Knowl 7 — Classification Performance on VisDA and DomainNet in Offline and Online Adaptation

Knowl 8 — Semantic Segmentation Performance on GTA5, SYNTHIA, and Dark-Zurich

Knowl 9 — Ablation of Selection Criteria and Loss Components in C-SFDA

Knowl 10 — Failure of Traditional Label Noise Learning Techniques in SFDA

Knowl 11 — Sensitivity to Initial Pseudo-Label Quality Under Severe Domain Shifts

References

Citation

BibTeX

Access the Paper