The loss function of a target pose estimation model based on a convolutional neural network (CNN) mostly uses the Euclidean distance between two points as the evaluation criterion. Although the loss function is simple in calculation and fast in operation, the training rules are not comprehensive enough and lack global understanding of the target. In this paper, a ComPoseNet model based on a combined loss function is proposed for pose estimation. The loss function in this model is based on spatial learning, and the two-point Euclidean distance, straight line, and straight line angle are used as training rules. Compared with the traditional loss function, this algorithm considers the spatial position of the target from the point, line, and angle, reducing the error between the estimated and the real poses so that the effect of the pose estimation is improved. Numerous experiments and analysis of LineMod data show that the algorithm has a higher convergence speed, greater accuracy, and smaller errors than the traditional algorithm operating for the same training times. The translation error is reduced by 7.407%, and the angle error is reduced by 6.968%.