Paper titled: Head Pose Estimation Based on 5D Rotation Representation [paper]

• Fig. Snapshots of 7 different views from one sequence in the CMU Panoptic val-set dataset.

Evaluation of the MAE for different methods on the CMU dataset

Method	R	RR	Yaw	Pitch	Roll	MAE
Narrow range: -90° < yaw < 90°
DirectMHP [Zhou et al., 2023]	✅	E	5.86	8.25	7.25	7.12
DirectMHP [Zhou et al., 2023]	❌	E	5.75	8.01	6.96	6.91
5DResNet (ours)	✅	5D	5.01	6.89	6.00	5.97
Full range: -180° < yaw < 180°
Cobo et al. [Cobo et al., 2024]	❌	E	-	-	-	10.47
Viet et al. [Viet et al., 2021]	❌	RM	9.55	11.29	8.32	9.72
Cobo et al. [Cobo et al., 2024]	❌	Q	-	-	-	9.32
DirectMHP [Zhou et al., 2023]	✅	E	7.38	8.56	7.47	7.80
DirectMHP [Zhou et al., 2023]	❌	E	7.32	8.54	7.35	7.74
WHENet [Zhou et al., 2020]	❌	E	8.51	7.67	6.78	7.65
Cobo et al. [Cobo et al., 2024]	❌	6D	-	-	-	7.45
5DResNet (ours)	✅	5D	5.96	7.68	6.28	6.64

RR: Rotation Representation. R: Retraining.

Dataset

• CMU Panoptic from here for the full range angles.

Training:

python3 train_5D_Resnet.py

Testing:

python3 test_Resnet.py

Citing

@inproceedings{algabri2024head,
  title={Head Pose Estimation Based on 5D Rotation Representation},
  author={Algabri, Redhwan and Lee, Sungon},
  booktitle={2024 IEEE Symposium on Wireless Technology \& Applications (ISWTA)},
  pages={195--199},
  year={2024},
  organization={IEEE}
}