電機資訊學院: 資訊工程學研究所指導教授: 王傑智陳彥廷Chen, Yen-TingYen-TingChen2017-03-032018-07-052017-03-032018-07-052014http://ntur.lib.ntu.edu.tw//handle/246246/275429大部分現有的影片標記系統(video annotation system)專注於標記影片中物體的行為(activity)，其他的系統則是致力於標記出影片中每個物體的位置甚至是物體的輪廓(object contour)。我們發現後者只利用定界框(bounding box)或是利用內插法幫助使用者標記一個物體在每個幀(frame)中對應的位置或輪廓，而其中只有一篇著作提及如何去找出被標記的物體之中的稠密對應關係(dense correspondence)。經過分析之後我們發現影片資料之稠密對應關係標記還有許多議題需要釐清。因此，我們發展了一個標記對應物體輪廓之中每個像素的對應關係的影片標記系統。 此外，由於標記整個影片中物體的細部輪廓以及稠密對應關係必須花費許多精神和時間，我們利用互動式分割(interactive segmentation)、光流法(optical flow)及邊緣檢測(edge detection)的結果讓使用者可以更容易觀察出兩個幀之間的明顯特徵對應關係(salient feature correspondence)。邊緣檢測的結果可以幫助使用者找出物體的細部輪廓或是物體局部的圖樣(local pattern)。我們要求使用者確認及修改演算法找出來的明顯特徵對應關係。而對於物體中沒有特徵的區域(textureless region)，我們將使用者標記在兩個相鄰幀的明顯特徵對應關係做非剛性對齊(non-rigid registration)來得到此區域的稠密對應關係。使用者只需要仔細的標記第一個幀的物體輪廓及明顯特徵然後再修正演算法錯誤的部分就可以將整個影片標記完成。實驗結果顯示我們的系統較適合用來標記非剛性的物體。There are a few existing annotation systems that aim to provide a platform for video annotation. Most of them focus on activity annotation while others concentrate on labeling individual objects. However, the latters focus on only labeling objects with bounding boxes or only using interpolation techniques to help user labeling. Moreover, only one of them try to find the dense correspondence inside the object contour. Issues of dense correspondences annotation across video frames are not well addressed yet. Inspired by this, a video annotation system that focuses on dense correspondences annotation inside the object contour is proposed in this work. In addition, since labeling detail object contour and dense correspondences across a whole video is a daunting task, we also minimize user''s effort by applying an interactive segmentation and tracking algorithm that utilizes information from optical flow and edges that helps the user easier to observe the salient feature correspondences between two video frames. Edges could help the user to find out the detail contour or local patterns of the object. The user is required to check and modify the salient feature correspondences obtained by the algorithm. Dense correspondences in the textureless region are extracted by a non-rigid registration algorithm from the salient feature correspondences verified by the user. The user only needs to label the first frame of the video and correct some minor errors in the subsequent frames for the whole video annotation. The result shows that the proposed framework is more suitable to label non-rigid objects.6882869 bytesapplication/pdf論文公開時間: 2015/2/3論文使用權限: 同意無償授權影片標記稠密對應明顯特徵對應非剛性對齊video annotationdense correspondencesalient feature correspondencenon-rigid registrationthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/275429/1/ntu-103-R01922116-1.pdf