qhjqhj00 multi-view-3d-pose-al-eval
Rethinking the Data Annotation Process for Multi-view 3D Pose Estimation with Active Learning and Self-Training — Qi Feng et al. (2021) (arXiv:2112.13709, 2021)
What this evaluates
Evaluates active learning strategies for multi-view 3D pose estimation by measuring annotation efficiency. It probes how well geometric consistency and self-training can reduce the number of required human annotations while maintaining low 3D keypoint error.
Datasets
- CMU Panoptic — total ?; splits: train (5008), val (891), test (771)
- InterHand2.6M — total ?; splits: train (12123), val (1900), test (1762)
Metrics
3D Mean Key Point Error (MKPE)(primary) — range: other- Average Euclidean distance between predicted and ground truth 3D keypoints across all joints and frames, measured in millimeters.
Input / output format
Input: Multi-view 2D images per frame, processed to predict 2D heatmaps for each camera view.
Output: 3D keypoint coordinates per frame, obtained via RANSAC triangulation of 2D keypoint predictions.
Scoring recipe
def compute_mkpe(pred_3d, gt_3d):
# pred_3d, gt_3d: (N_frames, N_joints, 3)
errors = np.linalg.norm(pred_3d - gt_3d, axis=-1)
return np.mean(errors)
Common pitfalls
- Extending single-view AL strategies to multi-view by simply aggregating per-frame uncertainty fails to account for geometric inconsistency, causing selection of frames that triangulate poorly.
- Sampling pseudo-labels from the full unlabeled set during self-training causes frame re-selection, leading to negative label drift over AL iterations.
Evidence (verbatim from paper)
For each experiment, we conduct 3 randomized trials, and report the average and variance for the 3D Mean Key Point Error (MKPE) in millimeter (mm).
Citation
@misc{feng2021rethinking,
title={Rethinking the Data Annotation Process for Multi-view 3D Pose Estimation with Active Learning and Self-Training},
author={Qi Feng et al. (2021)},
year={2021},
note={arXiv:2112.13709}
}
- arXiv: 2112.13709