MLPPose: Human Keypoint Localization via MLP-Mixer

Although existing methods have made great progress in human pose estimation, there are still a lot of challenging situations not well-handled, such as occluded limbs, invisible body parts or complex scenarios. In this work, we propose a novel approach called MLPPose, which combining the MLP-Mixer layers with the convolutional token embedding for human pose estimation. The MLP-Mixer layers are consisted of two types of MLP blocks, one concerns the global receptive field and the other mixes the channel feature at each location. This composition can not only obtain the association between different keypoints, but also efficiently capture the global dependency relationships between keypoints and scenes. Thus, it allows our model to efficiently locate the keypoints, despite that some of them are occluded, invisible or in complex scenarios. Meanwhile, it is able to simplify the progress of extracting the global dependency relationships compared to the attentional mechanism which is widely used in transformer models. Experiments show that our model achieves competitive results with state-of-the-art methods on the MS-COCO and MPII human pose estimation benchmarks. Moreover, our model is more lightweight and faster than other best performance methods.