陳炳宇臺灣大學：資訊管理學研究所鄭鎧尹Cheng, Kai-YinKai-YinCheng2007-11-262018-06-292007-11-262018-06-292006http://ntur.lib.ntu.edu.tw//handle/246246/54282在這篇論文裡面我們利用JAVA設計了一款小遊戲 ─ PhotoShoot。我們利用遊戲的方式，來大量蒐集ROI資料。當使用者玩我們的遊戲的時候，就順便幫我們標記了一張圖片上他們認為比較重要的區域。這些比較重要的區域，我們稱之為region-of-interest (ROI)。許多年來，很多的學者嘗試著發展出許多能自動在圖片中找到ROI的演算法。然而他們所各自發展出來的演算法卻難以互相比較彼此的正確率，這乃是因為缺乏了一個公正的比較基礎。本遊戲裡的照片，因為經過數千位玩家所標記，因此可以視為一個比較客觀而且公正的比較資料庫。有了這些公正的資料後，我們便可以客觀地比較不同演算法之間的效能差異，以及研究各種不同的ROI模型。而這些由玩家標記出來的結果，不但可以幫助電腦專家發展自動找尋ROI的演算法，也可以幫助心理學家及生理學家研究人類究竟是如何觀看照片，以及如何選取照片中有興趣的地方。In this thesis, we have developed a web game ─ PhotoShoot. When people play our game, they also help us to locate important areas in photos. These important areas are often called region-of-interest (ROI). Researchers have studied ROI for many years, and tried to retrieve ROI from an image using automatic methods. Lots of algorithms have been proposed, but it is very hard to compare their performance, since there is no common benchmark for comparison. Because our game has already been played by thousands of players, the results can help us to build a ROI ground truth model for each photo. With this ground truth database, we can easily compare these algorithms’ performance. Moreover, by observing the calculated ROI models, we also also draw some conclusions on the ROI prperties. We hope that our database and visualization results can benefit a lot to researchers of ROI, including computer scientists, psychologists and psychophysicists.1. Introduction 1 1.1. What is ROI? 1 1.2. What can ROI do? 2 1.3. Contributions 5 1.4. Orgnization 6 2. Related work 7 2.1. Non-automatic approach 7 2.2. Automatic approach 10 3. Game design 15 3.1. Goal 15 3.2. Inspiration 16 3.3. System architecture 17 3.4. Rules of the game 19 3.4.1. Targeting mode 20 3.4.2. Watching mode 21 3.4.3. Shooting mode 21 3.4.4. Game result 22 3.4.5. Scoring 23 3.4.6. Levels 24 3.4.7. High score ranking 25 3.4.8. Personal profile 26 3.5. Observations & some mechanism 26 3.5.1. Robot 26 3.5.2. Cheating 28 3.5.3. UI issues 29 3.5.4. Players’ feedback 30 4. ROI retrieving method 33 4.1. Build a candidate model 34 4.2. Vote the candidate 37 4.3. Temperature color visualization 41 4.4. ROI model refinement 43 4.5. Visualization web site 45 5. Results 47 5.1. ROI consistency rate 47 5.2. User strategy for photos without a clear ROI 50 5.3. Photo with no consistent ROI and center area 52 5.4. High contrast area 54 5.5. Text 55 5.6. Perspective view point 57 5.7. Photo with high level semantic meaning 59 5.8. Conclusions 61 6. Conclusion & future work 63 7. References 657081877 bytesapplication/pdfen-USROI遊戲JAVA人工智慧人類注意模型人類感知電腦視覺Region-of-Interest標記網路GameArtificial IntelligenceAttention ModelHuman PerceptionComputer VisionLabeingWebotherhttp://ntur.lib.ntu.edu.tw/bitstream/246246/54282/1/ntu-95-R93725053-1.pdf