使用人臉特徵點檢測與頭部姿勢估計的頭寬預測

Abstract

現今主要的應用都是透過相機捕捉人臉的圖像，再對該圖像進行偵測、辨識或利用類似擴增實境的方式對影像進行加工，鮮少能將影像的資訊再對應回真實世界。本論文把已發展出的人臉技術作為基礎，設計一套完整的流程架構，讓使用者可以透過單目攝像機拍攝的影像預測人臉頭部寬度的實際距離，並用得到的資訊作為線上眼鏡挑選的基礎。 我們透過頭部姿勢估計與人臉特徵點檢測得到人臉與五官的位置，根據特徵選擇演算法降低特徵維度，最後透過迴歸模型預測出頭寬，此外，我們搜集了一個包含多部自拍影片的資料集，並以此為基準來衡量距離換算的誤差，衡量該方法作為實際應用的可行性。

Nowadays, the main application is to capture the image of the human face through a camera and implement detection, recognition or use augmented reality of adding virtual objects to the image. However, most of them cannot be mapped back to reality. In this paper, we design a complete architecture that allows users to predict the real distance of human head width through a monocular camera based on state-of-the-art facial techniques. And use the information as a criterion for online eyeglasses purchasing. We obtain the position of facial landmarks through head pose estimation and facial landmark detection, reduce the feature dimension according to the feature selection algorithm, and finally predict the head width through the regression model. Besides, we collect a dataset that contains multiple selfie videos and use it as a benchmark to measure distance conversion and measure the feasibility of real applications of our methods.