This paper addresses the scale variation challenge in bottom-up human pose estimation. The authors present HigherHRNet, which leverages high-resolution feature pyramids to learn scale-aware representations. Built upon HRNet, the method combines "multi-resolution supervision for training and multi-resolution aggregation for inference" to generate high-quality multi-scale heatmaps. HigherHRNet achieves 70.5% AP on COCO test-dev, with a 2.5% AP improvement for medium persons, and 67.6% AP on CrowdPose, surpassing all top-down approaches.

Bottom-up human pose estimation detects all body joints first, then groups them into individual persons. Compared to top-down methods that first detect persons then estimate each person's pose, bottom-up approaches are inherently more efficient as their computation does not scale with the number of persons. However, bottom-up methods face a critical challenge: "the scale of persons in an image varies significantly," and single-resolution heatmaps struggle to accurately localize keypoints for both small and large persons simultaneously. Small persons require high-resolution heatmaps for precise localization, while large persons benefit from lower-resolution representations with larger receptive fields.

Existing bottom-up methods like OpenPose and Associative Embedding typically produce heatmaps at 1/4 of the input resolution, which loses fine-grained spatial information critical for small person detection. Simply increasing the heatmap resolution is computationally expensive and does not address the fundamental scale variation problem. The authors propose that generating multi-scale heatmaps through a high-resolution feature pyramid is the principled solution to simultaneously handle persons at different scales.

Top-down methods such as SimpleBaseline and HRNet achieve high accuracy by cropping each detected person and estimating pose on normalized patches, but their inference time scales linearly with the number of persons. Bottom-up methods including OpenPose (using Part Affinity Fields) and Associative Embedding (using tag-based grouping) are person-count-independent but historically lag behind in accuracy. Feature Pyramid Networks (FPN) have been successful in object detection for handling scale variation, but their application to pose estimation requires careful design due to the need for high-resolution spatial precision.

HigherHRNet builds upon HRNet as the backbone, which maintains high-resolution representations throughout the network via parallel multi-resolution branches with repeated information exchange. The key extension is a deconvolution-based upsampling module that generates feature maps at higher resolutions than HRNet's native output. Starting from HRNet's 1/4-resolution output, the module applies transposed convolutions to produce 1/2-resolution features, creating a high-resolution feature pyramid with heatmap predictions at multiple scales. This design enables the network to produce heatmaps where small persons are better represented at higher resolutions and large persons are captured at lower resolutions.

During training, the method applies heatmap supervision at each resolution level of the feature pyramid, with ground-truth heatmaps generated at corresponding scales. This multi-resolution supervision encourages each pyramid level to specialize in detecting persons of the appropriate scale. During inference, heatmap predictions from all resolution levels are aggregated by resizing them to a common resolution and averaging. This multi-resolution aggregation combines the complementary strengths of different scales — high-resolution precision for small persons and large-context awareness for large persons.

Experiments are conducted primarily on COCO and CrowdPose benchmarks. On COCO test-dev, HigherHRNet achieves 70.5% AP, setting a new state-of-the-art for bottom-up methods. The improvement is most pronounced for medium-sized persons (+2.5% AP), validating the scale-aware design. On CrowdPose, the method reaches 67.6% AP, surpassing top-down approaches that struggle with heavily overlapping persons where person detection is unreliable. Ablation studies confirm that both multi-resolution supervision and aggregation contribute to the final performance, with the high-resolution branch providing the most significant gains on small and medium persons.

HigherHRNet addresses the scale variation problem in bottom-up human pose estimation through high-resolution feature pyramids with multi-resolution supervision and aggregation. The method achieves state-of-the-art results on COCO and demonstrates particular strength in crowded scenes on CrowdPose. The scale-aware design principle provides a general framework for improving bottom-up pose estimation and potentially other dense prediction tasks requiring multi-scale reasoning.