AODPose:weakly Supervised Adversarial Learning With Ordinal Depth Supervision For 3D Human Pose Estimation

3D human pose estimation refers to estimating the 3D coordinates of the main joints from image or video.3D human pose estimation from a single image is a hot task in computer vision with many applications e.g.in human-computer interaction,action recognition,virtual reality,etc.Recently,most of methods for 3D human pose estimation from a single image firstly estimate 2D human pose from the input image and then estimate the 3D human pose from the 2D human pose.However,the estimation from the 2D pose to the 3D pose is ambiguous and is a morbid problem;Moreover,since it is difficult to perform 3D human pose annotation,the abundant 3D human pose training data is unavailable.The existing 3D human pose estimation methods still have many ways for improvement in terms of accuracy and generalization ability.In this paper,we study the estimation from 2D human pose to 3D human pose,focusing on how to effectively train the neural network to learn the 3D human pose through weak supervision,without 2D-3D correspondences.The main component of this paper is as follow:(1)A novel adversarial learning framework based on weakly-supervised way is proposed for 3D human pose estimation,which can enables the network to simultaneously learn 2D human pose,the ordinal depth between joints and the feature of human skeleton,effectively regressing 3D human pose from 2D human pose.(2)Based on adversarial learning framework,this paper introduces ordinal depth supervision of human pose to enhance the learning ability of generator.A novel loss function for ordinal depth supervision is also proposed to cooperatively train with critic,which make up for the critic's insufficient constraint on the ordinal depth.The experiments show that it can achieve more accuracy when combine the weak supervision of the ordinal depth and the critic.(3)Many experiments and visual analysis with Human3.6M and LSP datasets are carried out to prove the superiority of this method.Compared with existing method,the method proposed by this paper achieves more accuracy on the 3D human pose estimation.Besides,The proposed network have excellent generalization capabilities,which can generate a reasonable and accurate human pose for some complex pose that are not in the training data.