Multi-Scale Coordinate Classification for Human Pose Estimation Based on Feature Reassembly

Human pose estimation, as a highly significant research direction in computer vision, predominantly relies on heatmap-based methods and coordinate regression. Recently, coordinate classification-based approaches have gained popularity. In this paper, we propose a multi-scale coordinate classification network based on feature recombination. The network employs a pyramid-structured Transformer to acquire feature maps of varying resolutions. It leverages feature recombination to dynamically generate adaptive kernels for different feature maps, capturing location and keypoint sensitive features. Moreover, it aggregates the sampled (upsampled and downsampled) features to maximize the constraint on feature map quality. In addition, a coordinate-decoupled classification module is employed to classify 1D horizontal and vertical feature vectors, further reducing the network parameters and computational overhead. This approach avoids the expensive post-processing operations and quantization errors associated with heatmap methods, as well as the poor robustness issues of coordinate regression methods. The proposed model was evaluated on the COCO keypoint detection dataset and the MPII multi-person pose estimation dataset, and achieved satisfactory results. The performance of the Feature Recombination And Fusion Module (FRFM) is also demonstrated, emphasizing the significance of capturing detailed features for coordinate regression and classification methods.