Abstract
Traditional 2D human pose estimation (HPE) systems often struggle with keypoints that fall outside the image frame or are heavily occluded, typically resulting in uncalibrated heatmaps and "hallucinated" joint locations. We present ProbPose, a probabilistic framework that addresses these issues by predicting a calibrated presence probability and utilizing flexible probability maps. By moving beyond fixed Gaussian targets and incorporating a novel OKSLoss, ProbPose achieves state-of-the-art performance on standard benchmarks and shows significant robustness in handling out-of-image keypoints.
Introduction
Most top-down human pose estimation models operate on a cropped "activation window" centered on a detected person. While effective for centered subjects, this approach introduces significant failure modes. First, when a limb extends beyond the crop, models often mistakenly "snap" the hidden joint to a visible body part. Second, standard heatmaps use fixed Gaussian shapes that do not reflect the model's actual uncertainty. Finally, standard evaluation metrics like Object Keypoint Similarity (OKS) ignore missing keypoints, failing to penalize models for incorrect presence predictions.
Methodology
ProbPose introduces several key innovations to move towards a truly probabilistic representation of human pose:
Probability Maps and Sparsemax
Instead of regressing to a fixed Gaussian target using Mean Square Error (MSE), ProbPose employs Sparsemax to generate a probability map that sums to 1. This represents a true posterior distribution of the keypoint's location, allowing the model to represent complex, non-Gaussian uncertainty, such as bimodal distributions in ambiguous scenes.
Presence Probability ($p_p$)
A dedicated scalar output, the presence probability ($p_p$), is trained to predict whether a keypoint resides within the current activation window. This is distinct from standard confidence scores and is specifically optimized to handle partial crops and occlusions.
OKSLoss and Double-Probmap
The authors propose OKSLoss, a novel loss function formulated as an expected risk minimization problem that directly optimizes for the OKS metric. To handle keypoints far outside the image, a Double-Probmap strategy is used: a "global" map identifies the general vicinity, while an "expert" map provides high-precision localization within the standard window.
Experimental Results
The model was evaluated on the COCO dataset, as well as a new CropCOCO dataset designed to test out-of-image localization.
Quantitative Performance
On the standard COCO benchmark, ProbPose-s improved the mAP from 75.9 (ViTPose baseline) to 76.6. The gains were even more pronounced on CropCOCO, where mAP jumped from 72.7 to 81.7, demonstrating superior handling of truncated limbs. Furthermore, the model showed strong generalization on the OCHuman dataset (61.4 mAP), indicating that the probabilistic approach effectively manages heavy occlusion.
Qualitative Analysis
Qualitative results highlight ProbPose's ability to avoid the "hallucination" errors common in baseline models. By accurately predicting when a joint is out of frame and estimating its location beyond the crop boundaries, the model produces more anatomically plausible skeletons.
Conclusion
ProbPose demonstrates that a probabilistic framework significantly enhances the robustness and accuracy of 2D human pose estimation. By decoupling presence from localization and optimizing directly for OKS, the model overcomes the limitations of traditional heatmap-based approaches. Future work will explore the application of this framework to multi-person scenarios and 3D pose estimation, as well as addressing the noise in current dataset annotations near image borders.
Summary updated