用於直接從二維魚眼影像估計三維人體姿勢的一種深度學習方法

Abstract

在這份研究當中我們提出了一套基於深度學習的方法用來直接從二維魚眼影像估計人體關節在三維空間中的位置，這裡的二維魚眼影像是用一種以使用者自身為中心的視角去拍攝的。我們提出的方法之核心是一個基於Inception-v3所新設計的卷積類神經網路，特色是為魚眼影像調整而較大的卷積過濾器、訓練參數減量、長短期記憶、以及將擬人論的權重引入訓練網路時的損失函數。我們也進行了四類實驗來研究在該方法上使用不同訓練設定對測試結果的影響。這份研究的內容可以為發展出有著合理的資源使用量且更為複雜的電腦視覺深度學習網路提供經驗。

In this study, we propose a deep learning based method to directly estimate the human joint positions in 3D space from 2D fisheye images captured in an egocentric manner. The core of our method is a new design based on Inception-v3 convolutional neural network featuring the larger convolutional filter size, the reduction of parameters, the long short-term memory module, and the anthropomorphic weights on the training loss. We also conduct four groups of experiments to study the different effects upon the testing results when using different training settings of our work. The experience of our study can be helpful to develop more complicated deep learning network in a reasonable resource requirement to deal with the computer vision problems.