請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87623
標題: | 透過域解耦暨輔以源域引導取樣之對比學習於無監督領域自適應之行人重識別系統 Domain Disentanglement and Contrastive learning with Source-Guided Sampling for Unsupervised Domain Adaptation Person Re-identification |
作者: | 吳承軒 Cheng-Hsuan Wu |
指導教授: | 傅立成 Li-Chen Fu |
關鍵字: | 深度學習,行人重識別,領域適應,領域解耦,對比學習, Deep learning,Person re-identification,Domain adaptation,Domain disentanglement,Contrastive learning, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 近年來,因應大眾安全的需求與監控系統的大幅擴張,行人重識別在監控系統相關的研究中受到許多關注。在過去幾年中,基於監督式學習的行人重識別方法已經有很成熟的發展,並且在數個公開數據集中取得優異的成績。然而,因為現實環境的資料庫與訓練資料集間存在域差異,而且為每次佈署建立其有標註的資料集是十分耗費人力的,所以這些方法並不能夠輕鬆地移轉至現實生活中的應用。
為了能更適合於現實的系統佈署,基於無監督式學習的方法是一種有效的方案。在這無監督式學習中,主要可以分成兩個種類,分別是無監督式域適應以及完全無監督式學習。理論上,前者因為使用了額外的源資料集來訓練模型,所以表現應該要優於後者。然而目前的現況為完全無監督式學習的表現較優。這顯示現有的無監督式域適應方法無法可靠地將源域的知識傳遞到目標資料集上。為此,我們提出了一個基於無監督式域適應的方法,其中包括了域解耦網路以及源域引導之對比學習。域解耦網路會先將兩個資料集間的域差異縮小,之後則由源域引導之對比學習利用源域資料集的標註,引導在目標域上的訓練。得力於這兩個模塊,我們模型中的知識傳遞能夠順利地進行。 另外,因為人們對於隱私權的意識崛起,使的從公眾收集足夠的訓練資料變得越來越困難。這顯示為佈署環境建立一個新資料集的困難點不只在於為資料標註的勞力,也包括從公眾獲取資料的困難性。為了能夠更適合佈署於資料數受限的資料集,我們也進一步確保我們的模型在如此的情況下能夠有好的表現。除此之外,我們提出了一個小規模資料集的評估方法,以測試模型在上述情況下的學習能力。 為了證明此方法的有效性,我們在進行了充足的實驗,並證明我們的方法表現優於大多數現今表現最好的方法。 Recently, person re-identification (Re-ID) has raised a lot of attention in intelligent surveillance systems with significant research impact due to the urgent demand for public safety and the increasing number of surveillance cameras. In the past few years, fully supervised Person re-id methods have already been well developed and made tremendous performance on several public datasets. However, they cannot be easily applied to real-life applications because of the domain gap between real-world databases and training datasets as well as the high labor cost of constructing labeled datasets for every real-world deployment. To be easier for real-life system deployment, unsupervised learning is an effective solution. There are mainly two kinds of unsupervised person Re-ID solutions, which are unsupervised domain adaptation (UDA) and fully unsupervised learning (USL). Theoretically, the UDA methods should outperform the USL methods because they use additional source datasets. However, most of the existing state-of-the-art methods are USL-based until now. It shows that the existing UDA methods cannot transfer the knowledge from the source domain to the target dataset properly. Therefore, in this thesis we propose a UDA method that involves Domain Disentanglement Network (DD-Net) and Source-Guided Contrastive learning (SGCL). DD-Net first narrows down the domain gap between two datasets, and then SGCL utilizes the labeled source dataset as the clue to guide the training on the target domain. With the two modules, the knowledge transfer in our model can be completed successfully. Besides, as the awareness of right to privacy is on the rise, it becomes harder to collect sufficient training data from the public. Thence, the difficulty of constructing a new dataset for deployment not only arises from the labor cost of labeling but also because the raw data from the public are hard to come by. In order to be more suitable for the deployment with a limited-size dataset, we further make sure that our model performs well in such a situation. In addition, we propose a small-scale dataset evaluation protocol to examine the learning ability under the circumstance as mentioned above. In the conducted experiment, it is shown that the proposed method in this thesis outperforms the state-of-the-art methods. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87623 |
DOI: | 10.6342/NTU202210115 |
全文授權: | 同意授權(限校園內公開) |
電子全文公開日期: | 2025-12-07 |
顯示於系所單位: | 電機工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-1.pdf 目前未授權公開取用 | 8.2 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。