A simple and efficient channel MLP on token for human pose estimation

Human pose estimation is crucial to human-centered visual applications. Recently, transformer-based methods have achieved remarkable performance in human pose estimation. Transformers benefit from the self-attention mechanism, which calculates the correlation between keypoints and images. Furthermore, the multi-head attention mechanism further extends this idea, allowing the model to extract features from different attention. However, as the number of attention heads is increased, the model’s capacity to effectively process channel information becomes constrained. To overcome this limitation, a Channel MLP (CM) module is presented, which effectively improves the performance of TokenPose. The CM module consists of a channel attention mechanism integrated with a Multilayer Perceptron (MLP) block. The network evaluates the importance of each channel in this way, resulting in output features that contain more comprehensive information. The CM module enhances TokenPose’s ability to extract information effectively. Our model achieves 75.2 AP on COCO test-dev set and 90.4 PCKh@0.5 on MPII valid set while keeping similar parameters and computation as TokenPose.