不同於一般無監督域適應方法可使用源域資料，無源域適應語意分割目標是
在不參照源域資料下將源域模型遷移至目標域的資料分布。在沒有源域資料的參
考下，無源域適應語意分割模型容易導致不穩定的遷移，並只專注在主要類別而
忽略稀少類別。本文中，我們提出分佈轉移與適應性類平衡自學習架構，以解決
無源域適應語意分割問題。在分佈轉移階段，我們提出利用隱性特徵導正來縮減
源域與目標域的域間間隙。而在自學習階段，我們提出多重類負學習來降低預測
噪音，並且提出適應性類平衡閥值用以動態挑選類間偽標籤作為自學習標籤。在
街景分割資料集上的實驗結果表明，提出的方法明顯優於現有無源域適應語意分
割方法，甚至達到與可使用源域資料的無監督域適應方法相衡的表現。

Unsupervised Domain Adaptation (UDA) for semantic segmentation aims to transfer the knowledge learned from the source domain to the target domain. Unlike the source-available UDA setting, Source-Free Domain Adaptation (SFDA) has no access to the source data and rely solely on the well-trained source model for adaptation. Without the source data for reference, SDFA often leads to unstable adaptation and mostly focuses on common semantic classes. In this thesis, we propose a Distribution Transfer and Adaptive Class-balanced self-training (DTAC) framework to tackle the issues of SFDA for semantic segmentation. First, in the distribution transfer stage, we propose to narrow the domain gap by aligning the implicit feature characteristics of source model with the feature statistics of the target data. Next, in the self-training stage, we propose a multi-class negative learning method with adaptive thresholding to dynamically select robust pseudo labels for per-class self-supervision. Experimental results on urban scene benchmarks show that DTAC outperforms other SFDA baselines and even achieves competitive results with source-available UDA methods.

Contents
摘要 i
Abstract ii
Acknowledgements
1 Introduction 1
2 Related Work 4
2.1 Knowledge Distillation . . . 4
2.2 Self-Supervised Learning . . . 5
3 Method 6
3.1 Weight-Regularized Distribution Transfer . . . 6
3.2 Adaptive Class-Balanced Self-Training . . . 8
3.2.1 Multi-Class Negative Learning . . . 8
3.2.2 Adaptive Class-Balanced Thresholding . . . 9
4 Experiments 11
4.1 Datasets and Evaluation Metrics . . . 11
4.2 Implementation Details . . . 12
4.3 Comparison . . . 13
4.3.1 GTA5 → Cityscapes . . . 13
4.3.2 SYNTHIA → Cityscapes . . . 13
4.3.3 Visualization . . . 14
4.4 Ablation Study . . . 15
4.4.1 Effectiveness of Distribution Transfer . . . 16
4.4.2 Effectiveness of Multi-Class Negative Learning . . . 16
4.4.3 Effectiveness of Adaptive Class-Balanced Thresholding . . . 16
4.4.4 Effectiveness of DTAC over data-augmented baseline . . . 17
4.4.5 Hyperparameter selection . . . 18
5 Conclusion 21
References 22

[1] Y. Li, N. Wang, J. Shi, J. Liu, and X. Hou, “Revisiting batch normalization for practical domain adaptation,” arXiv preprint arXiv:1603.04779, 2016.
[2] M. Chen, H. Xue, and D. Cai, “Domain adaptation for semantic segmentation with maximum squares loss,” in Proc. ICCV, 2019.
[3] Y. Zou, Z. Yu, X. Liu, B. Kumar, and J. Wang, “Confidence regularized self-training,” in Proc. ICCV, 2019.
[4] T. Vu, H. Jain, M. Bucher, M. Cord, and P. Pérez, “Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation,” in Proc. CVPR, 2019.
[5] Y. Tsai, W. Hung, S. Schulter, K. Sohn, M. Yang, and M. Chandraker, “Learning to adapt structured output space for semantic segmentation,” in Proc. CVPR, 2018.
[6] N. Araslanov and S. Roth, “Self-supervised augmentation consistency for adapting semantic segmentation,” in Proc. CVPR, 2021.
[7] Y. Li, L. Yuan, and N. Vasconcelos, “Bidirectional learning for domain adaptation of semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6936–6945, 2019.
[8] Y. Zou, Z. Yu, B. Kumar, and J. Wang, “Unsupervised domain adaptation for semantic segmentation via class-balanced self-training,” in Proceedings of the European conference
on computer vision (ECCV), pp. 289–305, 2018.
[9] X. Luo, W. Chen, Y. Tan, C. Li, Y. He, and X. Jia, “Exploiting negative learning for implicit pseudo label rectification in source-free domain adaptive semantic segmentation,” arXiv
preprint arXiv:2106.12123, 2021.
[10] P. T. S and F. Fleuret, “Uncertainty reduction for model adaptation in semantic segmentation,” in Proc. CVPR, 2021.
[11] V. Prabhu, S. Khare, D. Kartik, and J. Hoffman, “S4t: Source-free domain adaptation for semantic segmentation via self-supervised selective self-training,” arXiv preprint arXiv:2107.10140, 2021.
[12] R. Lopes, S. Fenu, and T. Starner, “Data-free knowledge distillation for deep neural networks,” CoRR, vol. abs/1710.07535, 2017.
[13] H. Chen, Y. Wang, C. Xu, Z. Yang, C. Liu, B. Shi, C. Xu, C. Xu, and Q. Tian, “Data-free learning of student networks,” CoRR, vol. abs/1904.01186, 2019.
[14] P. Micaelli and A. Storkey, “Zero-shot knowledge transfer via adversarial belief matching,” in NeurIPS, 2019.
[15] Y. Liu, W. Zhang, and J. Wang, “Source-free domain adaptation for semantic segmentation,” in Proc. CVPR, 2021.
[16] R. Li, Q. Jiao, W. Cao, H. Wong, and S. Wu, “Model adaptation: Unsupervised domain adaptation without source data,” in Proc. CVPR, 2020.
[17] H. Yin, P. Molchanov, J. Alvarez, Z. Li, A. Mallya, D. Hoiem, N. Jha, and J. Kautz, “Dreaming to distill: Data-free knowledge transfer via deepinversion,” in Proc. CVPR, 2020.
[18] Z. Qiu, Y. Zhang, H. Lin, S. Niu, Y. Liu, Q. Du, and M. Tan, “Source-free domain adaptation via avatar prototype generation and adaptation,” in International Joint Conference on Artificial Intelligence, 2021.
[19] H. Xia, H. Zhao, and Z. Ding, “Adaptive adversarial network for source-free domain adaptation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9010–9019, 2021.
[20] J. Liang, D. Hu, Y. Wang, R. He, and J. Feng, “Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.
[21] S. Yang, J. van de Weijer, L. Herranz, S. Jui, et al., “Exploiting the intrinsic neighborhood structure for source-free domain adaptation,” Advances in Neural Information Processing
Systems, vol. 34, pp. 29393–29405, 2021.
[22] J. Huang, D. Guan, A. Xiao, and S. Lu, “Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data,” in NeurIPS, 2021.
[23] J. Kundu, A. Kulkarni, A. Singh, V. Jampani, and R. Babu, “Generalize then adapt: Source-free domain adaptive semantic segmentation,” in Proc. ICCV, 2021.
[24] Y. Kim, J. Yim, J. Yun, and J. Kim, “Nlnl: Negative learning for noisy labels,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 101–110, 2019.
[25] S. Richter, V. Vineet, S. Roth, and V. Koltun, “Playing for data: Ground truth from computer games,” in Proc. ECCV, 2016.
[26] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. Lopez, “The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,” in Proc. CVPR, 2016.
[27] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in Proc. CVPR, 2016.
[28] Y. Luo, L. Zheng, T. Guan, J. Yu, and Y. Yang, “Taking a closer look at domain shift: Category-level adversaries for semantics consistent domain adaptation,” in Proc. CVPR, 2019.
[29] J. Chang, Y.-T. Pang, and C.-T. Hsu, “Towards the target: Self-regularized progressive learning for unsupervised domain adaptation on semantic segmentation,” in Asian Conference on Pattern Recognition, pp. 299–313, Springer, 2022.
[30] L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. Yuille, “Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,” IEEE Trans. PAMI, vol. 40, no. 4, pp. 834–848, 2017.
[31] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. CVPR, 2016.