9D Rotation Representation-SVD Fusion with Deep Learning for Unconstrained Head Pose Estimation

Figure 1.

Figure 2.

Figure 3.

Figure 4.

Figure 5.

Comparison of the MAE between L2 and geodesic LOSS

	AFLW2000	BIWI	70/30 BIWI
Loss function	MAE	MAE	MAE
L2 Loss	3.90	3.92	2.71
Geodesic Loss	3.85	3.73	2.50

Comparisons with state-of-the-art methods on the AFLW2000 and BIWI dataset

AFLW2000					BIWI
Models	Yaw	Pitch	Roll	MAE	Yaw	Pitch	Roll	MAE
HopeNet[4]	6.40	6.53	5.39	6.11	4.54	5.15	3.37	4.36
FSA-Net[8]	4.50	6.08	4.64	5.07	4.64	5.61	3.57	4.61
HPE[6]	4.80	6.18	4.87	5.28	3.12	5.18	4.57	4.29
QuatNet[5]	3.97	5.62	3.92	4.50	2.94	5.49	4.01	4.15
WHENet[7]	5.11	6.24	4.92	5.42	3.99	4.39	3.06	3.81
TriNet[9]	4.04	5.77	4.20	4.67	4.11	4.76	3.05	3.97
FDN[10]	3.78	5.61	3.88	4.42	4.52	4.70	2.56	3.93
6DRepNet[18]	3.63	4.91	3.37	3.97	3.24	4.48	2.68	3.47
9D-EfficientNet	3.57	4.69	3.28	3.85	4.08	4.17	2.94	3.73

Comparison of MAE between ResNet and EfficientNetV2 backbone networks

	AFLW2000	BIWI	70/30 BIWI
Models	MAE	MAE	MAE
ResNetl8	4.37	3.70	2.64
EfficientNetV2-S	3.85	3.73	2.50

EfficientNetV2-S architecture

Stage	Operation	Stride	#Channels	#Layers
0	Conv3x3	2	24	1
1	Fused-MBConv1,3x3	1	24	2
2	Fused-MBConv4,3x3	2	48	4
3	Fused-MBConv4,3x3	2	64	4
4	MBConv4,3x3,SE0.25	2	128	6
5	MBConv6,3x3,SE0.25	1	160	9
6	MBConv6,3x3,SE0.25	2	256	15
7	Conv 1x1&Pooling&FC	-	1280	1

Euler error comparisons with state-of-the-art methods on the 70/30 BIWI dataset

	BIWI
Models	Yaw	Pitch	Roll	MAE
HopeNet[4]	3.29	3.39	3.00	3.23
FSA-Net[8]	2.89	4.29	3.60	3.60
TriNet[9]	2.93	3.04	2.44	2.80
FDN[10]	3.00	3.98	2.88	3.29
MDFNet[20]	2.99	3.68	2.99	3.22
DDD-Pose[21]	3.04	2.94	2.43	2.80
6DRepNet[18]	2.69	2.92	2.36	2.66
9D-EfficientNet	2.62	2.36	2.51	2.50

Comparison of parameters and FLOPs between 6DRepNet and our method

Models	Params	FLOPs
6DRepNet	43.752M	9.844G
9D-EfficientNet	20.189M	2.901G