ㅑTitle Page
Abstract
초록
Contents
Chapter 1. Introduction 15
1.1. Background and problem definition 15
1.2. Thesis outline 21
Chapter 2. Related works 22
2.1. Unsupervised domain adaptation 22
2.2. Source-free domain adaptation 25
2.3. Continual learning 28
2.4. Test-time adaptation 30
Chapter 3. Proposed method 33
3.1. Problem definition 34
3.2. Class-wise adaptation 38
3.3. Model-driven adaptive decision 41
Chapter 4. Evaluation and analysis of proposed method 45
4.1. Experimental settings 47
4.2. Experimental results 48
4.2.1. Result on CIFAR10C 48
4.2.2. Result on CIFAR100C 55
4.2.3. Result on ImageNet-C 58
4.3. t-SNE visualization 61
Chapter 5. Conclusion 62
Bibliography 64
Table 3.1. Performance variation based on the threshold value pth , the default threshold value for CoTTA is 0.92. the evaluation metric is error rate.[이미지참조] 35
Table 3.2. The performance varies depending on the frequency of pseudo- labeling[53], the evaluation metric is error rate. 41
Table 4.1. The classification results for adapting CIFAR10 to CIFAR10C in a continual setting. The results are presented in terms of the error rate. (In the case of EcoTTA, as the code is not provided and error rates for... 49
Table 4.2. Comparison of results in classifying the CIFAR10 to CIFAR10C adaptation task between the source pre-trained network method and the student... 53
Table 4.3. The results for CIFAR10 to CIFAR10C exhibit a gradual transition between the lowest and highest severity levels. These results are averaged across... 54
Table 4.4. The classification results for adapting CIFAR100 to CIFAR100C in a continual setting. The results are presented in terms of the error rate. (In the case of EcoTTA, as the code is not provided and error rates for... 57
Table 4.5. Classification results for ImageNet-to-ImageNet-C experiments con- ducted using the same experimental sequences as CIFAR benchmarks. The 58
Table 4.6. ImageNet to ImageNet-C experiments yielded an average error rate across ten diverse sequences of corruption types. 58
Table 4.7. Ablation study of our proposed method in the ImageNet to ImageNet- C task. The results are presented in terms of the error rate. The * denotes... 60
Figure 1.1. Domain shift problem 16
Figure 1.2. Unsupervised domain adaptation application (a) utilize domain adaptation in face recognition models where various conditions (b) adapting... 17
Figure 1.3. Comparison of domain adaptation scenarios: (a) The approach of unsupervised domain adaptation permits the utilization of both source and... 19
Figure 2.1. Domain adversarial neural networks [14] proposes a framework for unsupervised domain adaptation by introducing a domain classifier that encour-... 23
Figure 2.2. Source hypothesis transfer [24] introduces a method for transferring knowledge from a source domain to a target domain by leveraging source hy-... 25
Figure 2.3. Various continual learning methods: (a) regularizes weight approach, (b) utilizes memory replay approach, (c) utilizes parameter isolation method. 28
Figure 2.4. Pipeline of the Continual test-time adaptation (CoTTA) setting 30
Figure 3.1. Pseudo-labeling based on the decisions of a model updated on target data 36
Figure 3.2. Pipeline of the proposed class-wise adaptive test-time adaptation (CATTA) setting 44
Figure 4.1. 15 different corruption type 46
Figure 4.2. Corruption applied at different severity levels on clean images 46
Figure 4.3. t-SNE visualization (a) non-adapted (b) CoTTA adapted (c) CATTA adapted. Blue points are unseen source features and red are target features. 61