A Method of Constructing Fine-Grained Pose Evaluation Model

In the field of computer vision, human pose estimation (HPE) is an important direction. However, in some fields that require professional actions or motions, researchers will deal with poses with high diversity, strongly instantaneous actions and complex behaviors, and with problems such as lack of enough dataset, algorithm iteration lag, and poor robustness, it is difficult for conventional CV methods to recognize many kinds of poses well. Aiming to the problems above, we propose a method of training fine-grained pose estimation model and building pose dataset based multi-views in this paper. And during the work, we use the operation of filtering as optimization approach to improve the accuracy of results. By filtering data, those wrongly recognized key points can be corrected to some extent and the missing key points or poses can be estimated based on the continuity of motion. The filter method can also improve the smoothness of curve, so the pose sequences can be more fluent and conform the kinetics after the operation. Meanwhile, there sometimes may be errors about the results of 3D pose reconstruction based on 2D pose because of occlusions, changes of environment. In this situation, choosing 2D pose results from appropriate views instead of all the views can also improve the accuracy of final results. Moreover, for the environment with several persons, we finish the work of matching human cross-different views by combining with the method of predicting depths of root key points. From the experiments, the method can increase the average precision (AP) up to 2.2%, which demonstrate the feasibility and reusability of our method.