head-pose-estimation-adas-0001.md

head-pose-estimation-adas-0001

Use Case and High-Level Description

Head pose estimation network based on trimmed SqueezeNet_v1.1. Angle regression layers are convolutions (width: 128) + ReLU + batch norm + fully connected with one output.

Validation Dataset

Biwi Kinect Head Pose Database

Example

Specification

Metric	Value
Min face size	60x60 pixels (fixed)
Supported ranges	YAW [-90,90], PITCH [-70,70], ROLL [-70,70]
GFlops	0.1
MParams	1.9
Source framework	Caffe*

Accuracy

Angle	Mean ± standard deviation of absolute error
yaw	5.4 ± 4.4
pitch	5.5 ± 5.3
roll	4.6 ± 5.6

Inputs

Image in BGR format with size 60x60.

Outputs

Output layer names in Inference Engine format:

• angle_y_fc
• angle_p_fc
• angle_r_fc

Output layer names in Caffe* format:

• fc_y
• fc_p
• fc_r

Each output contains one float value that represents value in Tait-Bryan angles (yaw, pitсh or roll).

Legal Information

[*] Other names and brands may be claimed as the property of others.