Model
DAD-3DHeads: Results
The results of DAD-3DNet trained on DAD-3DHeads Dataset and evaluated on
3D Head Pose Estimation and 3D Face Shape Reconstruction benchmarks,
DAD-3DHeads Benchmark for 3D Head Estimation from dense annotations
suggest that dense supervision as provided in the dataset enables a holistic framework for 3D Head Analysis from images.
3D Head Pose Estimation results
DAD-3DNet largely outperforms the 3DMM estimation methods, and shows comparable performance to other SOTA methods.
BIWI Benchmark
MAE
|
Pitch MAE
|
Roll MAE
|
Yaw MAE
| |
|---|---|---|---|---|
HopeNet
|
4.90
|
6.61
|
3.27
|
4.81
|
Img2Pose
|
3.79
|
3.55
|
3.24
|
4.57
|
3DDFA-V2
|
8.81
|
12.08
|
7.54
|
6.80
|
RingNet
|
7.34
|
5.37
|
7.82
|
8.82
|
WHENet
|
3.81
|
4.39
|
3.06
|
3.99
|
DAD-3DNet
|
3.87
|
5.25
|
2.77
|
3.60
|
AFLW2000-3D Benchmark
Model
|
MAE
|
Pitch MAE
|
Roll MAE
|
Yaw MAE
|
|---|---|---|---|---|
HopeNet
|
6.16
|
6.56
|
5.44
|
6.47
|
RetinaNet
|
6.22
|
9.64
|
3.92
|
5.10
|
Img2Pose
|
3.91
|
5.03
|
3.28
|
3.43
|
SynergyNet
|
3.35
|
4.09
|
2.55
|
3.42
|
3DDFA-V2
|
7.56
|
8.48
|
9.89
|
4.30
|
RingNet
|
8.27
|
4.39
|
13.51
|
6.92
|
DAD-3DNet
|
3.63
|
4.73
|
3.19
|
2.98
|
3D Face Shape Reconstruction results
DAD-3DNet shows superior performance to the coarse 3D dense head alignment methods without explicitly disentangling Shape and Expression
NoW Benchmark (to be updated)
Model
|
Median (mm)
|
Mean (mm)
|
Std (mm)
|
|---|---|---|---|
3DDFA-V2
|
1.234
|
1.566
|
1.391
|
RingNet
|
1.207
|
1.535
|
1.306
|
DAD-3DNet
|
1.236
|
1.541
|
1.285
|
Feng et al. Benchmark
Model
|
3DRMSE
|
Median - HQ (mm)
|
Median - LQ (mm)
|
Mean - HQ (mm)
|
Mean - LQ (mm)
|
Std - HQ (mm)
|
Std - LQ (mm)
|
|---|---|---|---|---|---|---|---|
3DDFA-V2
|
2.998
|
1.500
|
1.779
|
1.942
|
2.350
|
1.704
|
2.149
|
RingNet
|
2.809
|
1.698
|
1.634
|
2.161
|
2.113
|
1.832
|
1.831
|
DAD-3DNet
|
2.718
|
1.523
|
1.634
|
1.957
|
2.096
|
1.691
|
1.808
|