Miroslav Purkrabek, Jiri Matas
2024-12-03
2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
10.1109/cvpr52734.2025.02526
13 citations
摘要
Current state-of-the-art Human Pose Estimation methods ignore out-of-image keypoints in both training and evaluation and use uncalibrated heatmaps as keypoint location representations. We propose ProbPose, which predicts for each keypoint: a calibrated probability of keypoint presence at each location in the activation window, the probability of being outside of it, and its predicted visibility. To address the lack of evaluation protocols for out-of-image keypoints, we introduce the CropCOCO dat...
The research addresses limitations in current human pose estimation (HPE) methods, which often overlook out-of-image keypoints and rely on uncalibrated heatmaps. This can lead to inaccurate keypoint localization, especially in challenging conditions like occlusions or cropped images.
The study introduces ProbPose, a model that predicts calibrated presence probabilities, keypoint locations, and visibility. It also introduces the CropCOCO dataset and Extended OKS (Ex-OKS) metric to evaluate out-of-image keypoints, using cropping-based data augmentation.
ProbPose demonstrates significant gains in out-of-image keypoint localization and improves in-image localization through data augmentation. Tested on COCO, CropCOCO, and OCHuman datasets, ProbPose shows enhanced robustness along bounding box edges and better flexibility in keypoint evaluation.