請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/5117
標題: | 透過穿戴式攝影機評估在社群中的視覺人類簽章認證 An Evaluation of Visual Human Signature Identification in Community via Wearable Camera |
作者: | Chia-Chin Tsao 曹嘉慶 |
指導教授: | 徐宏民(Winston H. Hsu) |
關鍵字: | 穿戴式裝置,人類屬性, Wearable device,Human attributes, |
出版年 : | 2014 |
學位: | 碩士 |
摘要: | 隨著穿戴式裝置的流行,我們逐漸能在不同的情境下接收到各種資訊。然而,囿於隱私權的規範,個人的資訊分享仍然是一個須要解決的問題。
我們提出一種不管從哪個方向/姿勢都能表達一個人的概念--人類視覺簽章 (VHS)。 使用者可以透過VHS將資訊散播於公開或是特定的社群中而不用顯示他們的身分。相對地,在社群中的人可以得知這些消息而不用知道這些人是誰。 這篇論文探討了一些可能對於不同角度跟姿勢具有不變性的樣式來建造VHS。 我們評測諸多有效於在不同角度辨識人的樣式在不同角度跟姿勢的情況下的表現,樣式包含了人的臉部外觀、視覺方塊、臉部屬性、衣服屬性。 我們還提出兩種用來融合多種樣式的方法--提高重要的維度以及加權融合,前者用來增加召回率後者用來增加準確率。 藉由同時考慮不同的樣式,我們提出的方法可以讓正面的VHS辨識達到51\%的辨識率,在最難的測試集底下達到23\%的辨識率。 為了完整的評測我們的成果,我們介紹一個包含從許多人從不同角度觀測以及不同姿勢拍攝的全新資料庫 -- 多角度名人個體資料庫 (MCID). 在這資料庫中擁有439位名人總共多於2000張從不同角度、不同服裝清晰的照片。據我們所知,這是截至目前為止能取得的資料庫中最大的。 With the increasing popularity of wearable devices, information is becoming easily available anywhere and anytime. However, personal information sharing still poses great challenges because of privacy issues. We propose an idea of Visual Human Signature (VHS) which can represent each person uniquely even captured in different views/poses by wearable camera. Users can post information to certain communities or public by their VHS without reveal their identification. Conversely, the community can find the information while detecting the corresponding VHS via wearable devices. The thesis explores some possible modalities to generate VHS invariant to different views and different poses. We evaluate the performance of multiple modalities including person's facial appearance, visual patches, facial attributes and clothing attributes which are effective for recognizing identity in different views. We also propose two methods to fuse the modalities -- emphasizing significant dimensions and weighted fusion; the former can improve the recall and the latter improve the precision. By jointly considering multiple modalities, our approach can achieve VHS recognition rate by 51\% in frontal images and 23\% in the most difficult dataset. To thoroughly evaluate our work, we introduce a new dataset for scenario of different view and clothing human retrieval called Multiview Celebrity Identity Dataset (MCID). The dataset contains more than 2,000 clarity images of 439 celebrities collected from web with different views and clothing. To the best of our knowledge, it is by far the largest publicly available multi-view and clothing dataset with identities. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/5117 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-103-1.pdf | 10.14 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。