<p>A limitation of many DETR-like methods in multi-person pose estimation is their insufficient utilization of multi-scale features, as features from the backbone are often passed directly to the transformer blocks. To address this, we present an end-to-end method that reconstructs the encoder’s output into multi-scale feature maps and enhances them for more robust predictions. First, we introduce a hybrid attention module that processes all scales of the reconstructed feature maps. By combining spatial and channel attention mechanisms, this module significantly improves the model’s ability to represent features, boosting its capacity to detect human bodies at various scales and adapt to occlusions and complex backgrounds. Second, our dedicated pose decoder refines the estimation process by replacing standard self-attention with distinct within-instance and cross-instance self-attention mechanisms. This design better captures the spatial relationships among keypoints within a single human instance while minimizing redundant interactions across different instances. Finally, we incorporate an effective query selection strategy to optimize the decoder’s initialization, which accelerates convergence and improves overall prediction accuracy. Experimental and qualitative results on the COCO and CrowdPose datasets validate the effectiveness of our method, demonstrating its robust and competitive performance in complex scenes.</p>