Uncovering the Blind Spots of OKS-mAP: A Unified Protocol for Occlusion, Joint-Angle, and Temporal Evaluation of 2D Pose Estimators

Authors: Yifan Fan

arXiv: 2611.12911 [cs.CV]

Submitted: 22 October 2026

Comments: 15 pages, 7 figures

Abstract

The dominant metric for 2D human pose estimation, OKS-mAP, has three blind spots: it does not distinguish visible from occluded keypoints, does not measure joint-angle error, and is silent on the temporal stability of tracking-free per-frame inference. We propose a supplementary protocol with three modules:

• (M1) OKS-mAP stratified by ground-truth visibility, with Δ_occ = mAP(v=1) − mAP(full), quantifying how much the aggregate metric masks occlusion-specific differences.
• (M2) Joint-angle error (RMSE/MAE/P95) plus knee-valgus detection consistency.
• (M3) Frame-to-frame displacement, median-filter residual jitter, and jump rate, before and after OneEuro smoothing.

We evaluated the protocol on four datasets (COCO val2017, CrowdPose, OCHuman, and PoseTrack18) and four models from two architecture families (YOLO11s-pose, an occlusion-domain fine-tuned variant, and ViTPose-B/L), under a common detector-free oracle protocol and an open-source harness.

Δ_occ is negative in every combination (−0.19 to −0.47): aggregate mAP overestimates occluded-scene performance in every tested combination. Joint-angle error is amplified in occluded layers, and angle metrics reveal differences that mAP cannot localize; for example, the smaller occlusion-trained model significantly outperforms a larger transformer on occluded elbows. On PoseTrack18 val, raw jitter matched between the two YOLO models, while pixel-domain OneEuro smoothing cut jump rates by 9–10% at no matched-scope AP cost. Residual jitter persists after detector removal, the transformer pair's jump rate is significantly lower, and smoothing benefits replicate across architectures. Stratified mAP tracks full mAP closely (sign-consistency 1.0, 12/12), serving as magnitude diagnosis rather than re-ranking; angle-metric difference vectors—especially for the knee—partially diverge, suggesting independent information. The protocol and harness are open-sourced, complementing rather than replacing OKS-mAP.

Key Contributions

1. Occlusion-stratified evaluation (M1): A visibility-stratified OKS-mAP analysis that exposes how aggregate scores mask occluded-scene performance gaps.
2. Joint-angle diagnostics (M2): Angle-error statistics and knee-valgus consistency that capture anatomical correctness beyond keypoint localization.
3. Temporal stability metrics (M3): Jitter and jump-rate measurements, including the effect of OneEuro smoothing, for tracking-free per-frame inference.
4. Open-source harness: A reproducible evaluation framework validated across four datasets and four models from two architecture families.



Metadata



• Subjects: Computer Vision and Pattern Recognition (cs.CV)
• Cite as: arXiv:2611.12911 [cs.CV]
• DOI: https://doi.org/10.48550/arXiv.2611.12911

via ArXiv CV