High-resolution representations are essential for position-sensitive vision problems, such as human pose estimation, semantic segmentation, and object detection. Existing state-of-the-art frameworks first encode the input image as a low-resolution representation through a subnetwork that is formed by connecting high-to-low resolution convolutions \emph{in series} (e.g., ResNet, VGGNet), and then recover the high-resolution representation from the encoded low-resolution representation. Instead, our proposed network, named as High-Resolution Network (HRNet), maintains high-resolution representations through the whole process. There are two key characteristics: (i) Connect the high-to-low resolution convolution streams \emph{in parallel}; (ii) Repeatedly exchange the information across resolutions. The benefit is that the resulting representation is semantically richer and spatially more precise. We show the superiority of the proposed HRNet in a wide range of applications, including human pose estimation, semantic segmentation, and object detection, suggesting that the HRNet is a stronger backbone for computer vision problems. All the codes are available at~{\url{https://github.com/HRNet}}.
核心贡献 · Key contributions
提出 HRNet，通过并行多分辨率卷积在整个过程中保持高分辨率表示。 Proposes HRNet that maintains high-resolution representations throughout the entire process via parallel multi-resolution convolutions.
引入重复多分辨率融合，在并行流之间交换信息，增强语义丰富性和空间精度。 Introduces repeated multi-resolution fusion to exchange information across parallel streams, enhancing semantic richness and spatial precision.
在人体姿态估计、语义分割和目标检测上，相比现有最优方法展现了优越性能。 Demonstrates superior performance on human pose estimation, semantic segmentation, and object detection compared to state-of-the-art methods.
提供两个版本：HRNetV1 用于姿态估计，HRNetV2 用于分割，HRNetV2p 用于检测，展示了多功能性。 Provides two versions: HRNetV1 for pose estimation and HRNetV2 for segmentation, with HRNetV2p for detection, showing versatility.
建立了多分辨率融合与常规卷积之间的联系，为探索所有分辨率提供了依据。 Establishes connection between multi-resolution fusion and regular convolution, justifying exploration of all resolutions.
在多个基准上以相当的模型大小和更低的计算复杂度取得了最优结果。 Achieves state-of-the-art results with comparable model size and lower computation complexity on multiple benchmarks.
局限 · Limitations
目标检测训练时的内存成本略高于现有最优方法。 Memory cost during training for object detection is slightly larger than state-of-the-art methods.
在 PyTorch 上姿态估计的训练时间稍长，但在支持静态图的 MXNet 上相当。 Training time for pose estimation is a little larger on PyTorch, though comparable on MXNet with static graph.
在 Cityscapes 分割中，HRNetV2 相比 HRNetV1 对大模型的性能提升较小。 Performance gain from HRNetV2 over HRNetV1 is minor for large models in Cityscapes segmentation.
研究未探索高于输入 1/4 的分辨率，留下了进一步改进的潜力。 The study does not explore resolutions higher than 1/4 of input, leaving potential for further improvement.
与 ASPP 或 PPM 结合未改善 Cityscapes 分割，限制了某些融合策略。 Combination with ASPP or PPM did not improve Cityscapes segmentation, limiting certain fusion strategies.