具有生產力的遊戲正確率分析-以建築地標標註為例

Abstract

儘管電腦科學已有了相當長遠的發展，還是存在著一些問題是人們可以輕易解決，但是電腦卻沒有辦法的，像是圖片辨識(image recognition)和圖片標註(image annotation)。而人力計算(human computation)是專門在研究這類問題的一個新的研究領域，本篇論文即是利用人力計算的方法，讓人們來幫助解決這些電腦沒有辦法有效解決的問題。 在這裡我們提出了一個兩人網路線上遊戲Image Hunter來幫助完成網路上建築地標圖片的標註(landmark image annotation)。除此之外，我們著重在分析遊戲中收集到的資料的品質分析。我們提出信任值估算(confidence evaluation)的方法取代了次數累積(times accumulation)的方法來評估資料收集的正確性。我們請來了28位受測者來進行實驗，實驗結果驗證了遊戲機制的合理性，以及信任值估算的方法可以有效地提升收集到的資料的正確性。

Despite impressive advancement in computer technology, there are still some problems that humans can solve efficiently but current computer programs can not. Image recognition and annotation are examples. Human computation is a new research area that focuses on this kind of problems. This thesis aims to explore the power of human computation and shows how humans could help solve problems that are hard for computers. We propose a two-player online human computation game, Image Hunter, to achieve the task of annotating for a collection of landmark images on the Internet. Moreover, we address on the quality analysis for the data collected by the game. We propose confidence evaluation instead of times accumulation to estimate the accuracy of the data. Experiments involving 28 players have been conducted. The preliminary results demonstrate that the game mechanism is reasonable and with confidence evaluation mechanism the accuracy improves effectively.