This is an official pytorch implementation of Deep High-Resolution Representation Learning for Human Pose Estimation. In this work, we are interested in the human pose estimation problem with a focus on learning reliable high-resolution representations. Most existing methods recover high-resolution representations from low-resolution representations produced by a high-to-low resolution network. Instead, our proposed network maintains high-resolution representations through the whole process. We start from a high-resolution subnetwork as the first stage, gradually add high-to-low resolution subnetworks one by one to form more stages, and connect the mutli-resolution subnetworks in parallel. We conduct repeated multi-scale fusions such that each of the high-to-low resolution representations receives information from other parallel representations over and over, leading to rich high-resolution representations. As a result, the predicted keypoint heatmap is potentially more accurate and spatially more precise. We empirically demonstrate the effectiveness of our network through the superior pose estimation results over two benchmark datasets: the COCO keypoint detection dataset and the MPII Human Pose dataset. The code and models have been publicly available at \url{https://github.com/leoxiaobin/deep-high-resolution-net.pytorch}.
核心贡献 · Key contributions
提出高分辨率网络（HRNet），在整个过程中保持高分辨率表示。 Proposes High-Resolution Net (HRNet) that maintains high-resolution representations throughout the whole process.
并行连接高到低分辨率子网络，而非串联，实现空间精确的热图预测。 Connects high-to-low resolution subnetworks in parallel instead of series, enabling spatially precise heatmap prediction.
引入跨并行子网络的重复多尺度融合，以丰富高分辨率表示。 Introduces repeated multi-scale fusion across parallel subnetworks to enrich high-resolution representations.
在 COCO 关键点检测、MPII 人体姿态和 PoseTrack 数据集上取得最先进结果。 Achieves state-of-the-art results on COCO keypoint detection, MPII Human Pose, and PoseTrack datasets.
与先前方法相比，以更低的计算成本和模型大小展现出卓越效率。 Demonstrates superior efficiency with lower computational cost and model size compared to prior methods.
通过多尺度融合和分辨率保持的消融研究验证了有效性。 Validates effectiveness through ablation studies on multi-scale fusion and resolution maintenance.
局限 · Limitations
专注于单人姿态估计；多人场景依赖外部人体检测器。 Focuses on single-person pose estimation; multi-person scenarios rely on external person detectors.
在 MPII 数据集上性能饱和；更大模型未提升结果。 Performance on MPII dataset saturates; larger model does not improve results.
需要针对不同任务仔细调整网络深度和宽度。 Requires careful tuning of network depth and width for different tasks.
多尺度融合在训练期间增加内存和计算量。 Multi-scale fusion increases memory and computation during training.
评估仅限于人体姿态估计；对其他密集预测任务的泛化性未充分探索。 Evaluation limited to human pose estimation; generalization to other dense prediction tasks not fully explored.
论文章节 · Sections（共 12）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 方法3 Approach
4.1 COCO 关键点检测4.1 COCO Keypoint Detection
4.2 MPII 人体姿态估计4.2 MPII Human Pose Estimation
4.3 在姿态追踪中的应用4.3 Application to Pose Tracking
4.4 消融研究4.4 Ablation Study
5 结论与未来工作5 Conclusion and Future Works
MPII 验证集上的结果Results on the MPII Validation Set
PoseTrack 数据集上的更多结果More Results on the PoseTrack Dataset
ImageNet 验证集上的结果Results on the ImageNet Validation Set