Angle Estimation Performance(Paper Layout)

Table 1. Angle Error: RAW vs USED (Canny post-processing)

Model	Params	Type	Mean Error			Robust Statistics
Model	Params	Type	avg err 012 (°)	avg err 123 (°)	p1 err (px)	Median (°)	MAD (°)	p95 (°)
YOLOv8n-pose	3.30 M	RAW	9.67	11.27	8.37	7.79	5.44	32.85
YOLOv8n-pose	3.30 M	USED	7.86	8.93	7.37	5.77	3.93	24.78
YOLOv8s-pose	11.63 M	RAW	11.02	14.23	10.36	6.96	4.84	36.61
YOLOv8s-pose	11.63 M	USED	11.04	14.54	9.68	6.96	4.17	43.18
YOLO11n-pose	2.91 M	RAW	10.45	10.10	9.37	6.14	4.51	30.85
YOLO11n-pose	2.91 M	USED	11.09	10.98	9.75	6.13	4.55	42.30
YOLO11s-pose	9.95 M	RAW	17.99	16.72	14.39	14.40	9.22	41.68
YOLO11s-pose	9.95 M	USED	16.99	16.61	14.20	13.79	8.88	44.90

Table 2. Detection Quality Metrics (best_angle.pt)

Model	Box Metrics				Pose Metrics
Model	Precision	Recall	mAP50	mAP50-95	Precision	Recall	mAP50	mAP50-95
YOLOv8n-pose	0.885	0.885	0.921	0.623	1.000	1.000	0.995	0.891
YOLOv8s-pose	0.996	0.846	0.934	0.604	1.000	0.962	0.981	0.868
YOLO11n-pose	0.956	0.842	0.912	0.617	1.000	0.923	0.961	0.866
YOLO11s-pose	0.962	0.962	0.978	0.421	1.000	1.000	0.995	0.812

Figures

Figure 1. Mean angle error comparison between RAW and USED outputs (lower is better).

Figure 2. Pose detection quality metrics (higher is better).

Figure 3. Robust error statistics based on RAW outputs.

Figure 4. Radar plot of normalized USED error metrics (outer ring indicates higher error; inner ring indicates lower error).

Figure 5. Radar plot of normalized detection quality metrics (outer ring indicates higher score).

USED = post-processed with Canny edge snap. Figure 4 uses USED error metrics with intuitive direction (inner=lower error, outer=higher error). Figure 5 uses box/pose metrics from best_angle.pt. All reported metrics are based on models trained with offline augmentation.