Introduction: The Localization Gap
In 2D human pose estimation, researchers typically choose between top-down methods (detecting a person box, then joints) and bottom-up methods (detecting all joints, then grouping). While top-down methods are currently more accurate, they are computationally heavy due to their reliance on external detectors.
Bottom-up methods are faster but suffer from a "localization gap." Specifically, "center-offset" approaches—which predict a person's center and then regress offsets to every joint—often struggle with distant joints like wrists or ankles due to the high structural variance of the human body. The authors of LOGO-CAP investigate whether the information for accurate localization is present in the network but simply missed by the regression head.
The LOGO-CAP Framework
LOGO-CAP (LOcal-GlObal Contextual Adaptation for human Pose) addresses the localization gap through a three-step process designed to "refocus" misplaced joints:
- Initial Estimation: A backbone (HRNet) predicts person centers and initial offset vectors to keypoints, similar to standard models like CenterNet.
- Keypoint Expansion (Lifting): Instead of relying on a single $(x, y)$ coordinate, the model "lifts" each keypoint into a Keypoint Expansion Map (KEM), a small $11 \times 11$ local window.
- Contextual Adaptation: A Convolutional Message Passing (CMP) module analyzes these local windows to understand structural relationships. It generates Keypoint Attraction Maps (KAMs), which act as dynamic convolutional kernels. These kernels are applied to global keypoint heatmaps to "attract" the initially misplaced keypoints toward their true, high-confidence locations.
This approach is end-to-end trainable and avoids the complex, hand-crafted grouping algorithms used in older bottom-up methods.
Experimental Results
The researchers evaluated LOGO-CAP on the COCO Keypoint Challenge and the OCHuman dataset, which focuses on occluded and crowded scenes.
Key Insights
A critical "Aha!" moment occurred when the authors discovered that performing a local search in an $11 \times 11$ window around a vanilla model's predictions could theoretically jump accuracy from 60.1 AP to 88.9 AP. This proves the network "knows" where the joints are; the regression head is simply not precise enough. LOGO-CAP uses dynamic refocusing to bridge this gap.
Performance and Efficiency
On the COCO val-2017 dataset, LOGO-CAP achieved 70.0 AP, a 9.9-point increase over the baseline. It outperformed other leading bottom-up methods like DEKR and HigherHRNet. Furthermore, using an HRNet-W32 backbone, the model runs at 20.7 FPS on a TITAN RTX GPU, making it viable for near real-time applications.
Conclusion
LOGO-CAP significantly closes the gap between bottom-up and top-down pose estimation. While a 19-point difference remains between its current performance and the empirical upper bound of 88.9 AP, the success of dynamic refocusing suggests that even better mechanisms are possible. Future work may involve replacing the simple CMP module with Transformers to better capture complex, long-range structural dependencies.
Summary updated