- 學位論文
基於人臉相似圖傳遞人名之人物搜尋優化
Boosting People Search in Large-Scale User-Contributed Videos by Name Propagation over Face Graph
若您是本文的作者，可授權文章由華藝線上圖書館中協助推廣。
摘要
近年來隨著影片分享網站如Youtube、Google Video的興起，影片觀看習慣有了重大的改變，使用者透過在網站上隨意瀏覽或是鍵入特定的關鍵字觀看影片。一般來說，人物常是影片中的主角，因此在這類使用者提供影片上的人物搜尋也成為一個重要課題。現今類似網站上的影片搜尋都是基於直接搜尋影片周圍的相關文字資訊，而這些文字資訊因為使用者下標籤的行為模式而導致其有缺失或錯誤。我們提出了一個利用在人臉相似圖上傳遞人名以優化在這類影片上人物搜尋的效果。針對每部影片，我們透過影片在時間上的連續性質，擷取出較佳的正面人臉：而為了減少在幾乎相同的人臉上的不必要計算，我們也會針對每部影片中的所有人臉做分群(Clustering)的動作，經由此動作產生的人臉群集不但可以初略區分出影片中出現的人物並且透過使用群集為單位做之後的處理也能節省大量計算。利用由所有影片所產生的人臉群集，可以建立起一個人臉相似圖，進而在上面進行跨影片的人名傳遞。而為了處理大規模資料量，我們採用了數種基於雜湊(Hash)的方法建立人臉相似圖。實驗結果驗證我們所提出的方法確實可以優化人物搜尋的結果，找到更多原本沒有標籤的目標人物影片片段；另外，在運算時間方面，經由雜湊建立人臉相似圖的方法，所需時間只有暴力法(Brute force)的四百二十三倍分之一，擁有應用在大規模資料上的能力。
並列摘要
People search is one of the important needs for locating specific people (celebrities) in the explosive user-contributed videos. Current methods solely rely on (noisy) text (e.g., user-provided tags, descriptions) for the video. Meanwhile, we are also interested in finding specific segments that the target people appear rather than have to manually browse for each candidate videos. We propose name propagation through the face graph constructed in an unsupervised manner. In order to reduce unnecessary computation for duplicate face images and identifying salient faces for robust matching, we conduct local clustering within each video by affinity propagation with adaptive cluster number and selected exemplar faces for each cluster. And a similarity graph for face clusters across videos is built by measuring the visual similarities between the exemplar faces. Hash-based methods are applied to reduce the computation time and memory space for building this graph. The candidate names (with likelihood) in each cluster is then determined from their neighboring face clusters and weighted by the similarities. Evaluating in over five thousand YouTube videos, the experiment confirms that the proposed method is robust to noisy (or missing) tags commonly observed in user-contributed videos, saliently outperform the text-based method, and even can locate video segments where people of interest appear.
參考文獻
[1] T. Ahonen, A. Hadid, and M. Pietikainen. Face recognition with local binary patterns. In ECCV, 2004.
[2] T. Ahonen, A. Hadid, and M. Pietikainen. Face description with local binary patterns: Application to face recognition. In PAMI, 2006.
[4] T. L. Berg, E. C. Berg, J. Edwards, and D. A. Forsyth. Who is in the picture. In NIPS, 2006.
[6] H. Cevikalp and B. Triggs. Face reecognition based on image sets. In CVPR, 2010.
[7] Charikar and M. S. Similarity estimation techniques from rounding algorithms. In STOC, 2002.