Search papers, labs, and topics across Lattice.
This paper enhances YOLO-Pose models by introducing a post-hoc probabilistic extension that quantifies spatial uncertainty in keypoint localization through calibrated bivariate predictive distributions. By training additional probabilistic heads to predict a $2\times2$ dispersion matrix for each keypoint, the authors achieve improved reliability rankings and effective uncertainty-based pruning of unreliable keypoints. Experiments demonstrate that the proposed method significantly enhances keypoint-level precision and supports applications such as uncertainty-aware aircraft position estimation during landing.
Calibrated uncertainty estimates from YOLO-Pose models enable reliable keypoint ranking and effective pruning, transforming how we assess localization confidence.
YOLO-Pose models provide efficient keypoint localization, but do not quantify the associated spatial uncertainty. We introduce a lightweight post-hoc probabilistic extension that augments a trained YOLO-Pose model with calibrated bivariate predictive distributions over keypoint locations, centered at the model's original predictions. Concretely, we train additional probabilistic heads with an importance-weighted negative log-likelihood to predict an input-dependent $2\times2$ dispersion matrix for each keypoint, followed by Gaussian calibration for broad downstream compatibility or Student-$t$ calibration for distributional fidelity. Complementing this, we propose an evaluation protocol that combines a suite of distributional calibration diagnostics with average keypoint precision (AKP), a keypoint-level extension of the COCO AP protocol for assessing reliability rankings. Experiments on COCO show that the learned uncertainty estimates enable effective keypoint-level reliability ranking, Student-$t$ calibration best captures the empirical residual distribution, and uncertainty-based pruning removes unreliable keypoints. A central application-level demonstration is vision-based aircraft landing, where calibrated covariances for runway keypoints support uncertainty-aware aircraft position estimation and downstream sensor fusion.
