Deep High-Resolution Representation Learning for Visual Recognition

HRNet is a backbone architecture for position-sensitive computer vision tasks such as human pose estimation, semantic segmentation, and object detection. Whereas existing state-of-the-art frameworks first encode an image into a low-resolution representation via serially connected high-to-low resolution convolutions and then attempt to recover high resolution, HRNet instead maintains high-resolution representations through the entire process. Its two key ideas are connecting high-to-low resolution convolution streams in parallel and repeatedly exchanging information across those resolutions.

Because high resolution is preserved and information is fused across scales, the resulting representation is both semantically richer and spatially more precise. The authors show HRNet's superiority across a wide range of applications including pose estimation, semantic segmentation, and object detection, demonstrating that it is a stronger general-purpose backbone for computer vision, with all code released publicly.