ContQuat

ContQuat: Continuous Quaternion Representation for Head Pose Estimation here

• Fig. Framework of the proposed ContQuat model.

Results visualization

• Fig. Snapshots of different views.

Our results

• Table 1: MAE values for the CMU dataset obtained using different methods for Narrow-range angles: −90◦ < yaw < 90◦.

Method	Retrain	Rep.	Yaw ↓	Pitch ↓	Roll ↓	MAE ↓
DirectMHP [1]	✅	E	5.86	8.25	7.25	7.12
DirectMHP [1]	❌	E	5.75	8.01	6.96	6.91
6DRepNet [2]	✅	6D	5.20	7.22	6.00	6.14
ContQuat (Ours)	✅	CQ	4.64	6.84	5.54	5.67

• Table 1: MAE values for the CMU dataset obtained using different methods for Full-range angles: −180◦ < yaw < 180◦.

Method	Retrain	Rep.	Yaw ↓	Pitch ↓	Roll ↓	MAE ↓
Viet et al. [1]	❌	RM	9.55	11.29	8.32	9.72
WHENet [2]	❌	E	8.51	7.67	6.78	7.65
DirectMHP [3]	✅	E	7.38	8.56	7.47	7.80
DirectMHP [3]	❌	E	7.32	8.54	7.35	7.74
Cobo et al. [4]	❌	6D	-	-	-	7.45
6DRepNet [5]	✅	6D	5.89	7.76	6.39	6.68
ContQuat (Ours)	✅	CQ	5.36	7.49	6.14	6.33

• Table 2: MAE values for the BIWI and AFLW2000 datasets using different methods.

Method	Retrain	Rep.	AFLW2000 Yaw ↓	AFLW2000 Pitch ↓	AFLW2000 Roll ↓	AFLW2000 MAE ↓	BIWI Yaw ↓	BIWI Pitch ↓	BIWI Roll ↓	BIWI MAE ↓
FSA-Net [1]	❌	E	4.50	6.08	4.64	5.07	4.27	4.96	2.76	4.00
WHENet [2]	❌	E	4.44	5.75	4.31	4.83	3.60	4.10	2.73	3.48
TokenHPE [3]	❌	RM	4.36	5.54	4.08	4.66	3.95	4.51	2.71	3.72
QuatNet [4]	❌	Q	3.97	5.62	3.92	4.50	4.01	5.49	2.94	4.15
LSR [5]	❌	E	4.26	5.27	3.89	4.47	4.29	3.09	3.18	3.52
MFDNet [6]	❌	RM	4.30	5.16	3.69	4.38	3.40	4.68	2.77	3.62
SRNet [7]	❌	Q	3.75	5.10	3.46	4.10	3.81	4.36	2.77	3.65
DirectMHP [8]	❌	E	2.99	5.35	3.77	4.04	3.57	5.47	4.02	4.35
Li et al. [9]	❌	RM	3.36	5.05	3.56	3.99	3.59	3.94	2.68	3.40
6DRepNet [10]	❌	6D	3.63	4.91	3.37	3.97	3.24	4.48	2.68	3.47
FSA-Net [1]	✅	E	5.41	6.82	5.42	5.88	4.74	5.32	3.26	4.44
LSR [5]	✅	E	5.96	6.45	4.19	5.77	4.43	3.98	3.52	3.98
6DRepNet [10]	✅	6D	3.50	4.81	3.47	3.93	3.79	4.53	2.89	3.74
ContQuat (Ours)	✅	CQ	3.36	4.69	3.31	3.79	3.91	4.43	2.69	3.68

Datasets

• CMU Panoptic from here for the full range angles.

• 300W-LP, and AFLW2000 from here for the narrow range angles.

• BIWI from here for the narrow range angles.

For Training:

If you only need to change the pre-trained RepVGG model 'RepVGG-B1g2-train.pth' please see here and save it in the root directory.

python3 train.py

After training is done. Next step.

For Deploy models:

For reparameterization, the trained models into inference models use the convert script.

python3 convert.py input-model.tar output-model.pth

After converting the training model into an inference model. Then, you can test your model.

For Testing:

python3 test.py

If you are too lazy to do the above steps, you can download the pre-trained RepVGG model to run your application. 'CMU_best.pth here' for the full range angles or '300_best.pth here' for narrow range angles.

Citing

@article{abdu2026contquat,
  title={ContQuat: Continuous quaternion representation for head pose estimation},
  author={Abdu, Ahmed and Bae, Ji-Hun and Lee, Sungon and Algabri, Redhwan},
  journal={Information Sciences},
  pages={123621},
  year={2026},
  publisher={Elsevier}
}