MixPose: 3D Human Pose Estimation with Mixed Encoder

The fusion of spatio-temporal information is crucial for 3D human pose estimation in video. Existing methods usually extract temporal information from the spatially encoded poses, which may lead to limited spatio-temporal information interaction. To address this issue, we propose MixPose, a novel network for 3D human pose estimation with mixed encoder in videos. We introduce independent mixed encoders to fuse spatio-temporal information in the sequence, and augment the perception of each point with global information using an attention module. We evaluate MixPose on two public datasets, Human3.6M and HumanEva, experiment results show that MixPose outperforms other state-of-the-art methods in specific scenarios.