Selecting suitable robot hand poses for grasping objects of various shapes and sizes remains a challenge in the control of humanoid robot hands. This study proposes a method to generate a robot hand pose from a self-built data set of 22 finger joint angles, which can be used as input for the robot hand forward kinematics model to grasp an object. The dataset is constructed from the shape and size properties of more than 4000 object images, combined with the human hand joint angles collected by a data glove while holding the objects. Initially, the shapes and sizes of objects pictured by a depth camera are identified using a Convolutional Neural Network (CNN) algorithm. A data-collecting glove equipped with 13 IMU sensors is then employed to build a dataset of hand poses along with corresponding labels for the recognized object's shape and size. Correspondingly, when a new object is pictured by the camera, its shape and size are matched into the prebuilt data set of hand poses to generate the most suitable hand pose for robot grasping control.