半合成数据集增强

本研究创建了一种半合成数据集增强方法，用于特定应用场景下的视线估计。该数据集通过生成具有纹理的三维面部网格，并从与应用相关的特定位置和方向的虚拟相机渲染训练图像来扩展现有数据集。此方法特别适用于辅助机器人或市场研究等应用，其中视线点可能不在相机原点附近。数据集创建过程中，首先对每个数据集样本的面部进行对齐，然后生成带纹理的面部网格，并应用逆姿态变换来校正视线方向。该数据集的应用领域主要集中在提高辅助机器人任务中的视线估计准确性，特别是在自动化抓取等任务中。

This study proposes a semi-synthetic dataset augmentation method for gaze estimation in specific application scenarios. This method expands existing datasets by generating textured 3D facial meshes and rendering training images from virtual cameras positioned at specific locations and orientations relevant to the target applications. This approach is particularly suitable for applications such as assistive robotics or market research, where gaze points may not be located near the camera origin. During the dataset creation process, the facial region of each dataset sample is first aligned, followed by the generation of textured facial meshes and the application of inverse pose transformation to correct the gaze direction. The primary application scope of this dataset focuses on improving the accuracy of gaze estimation in assistive robotic tasks, especially in scenarios like automated grasping.