Head pose estimation

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

Model Details

Model Description

Head pose estimation model is a lightweight, simple, and handmade convolutional neural network. The model works on cropped-face images and predicts how the face is oriented. This let us detect where the specific person is looking. It outputs 3 values yaw, pitch and roll.

Developed by: OpenVINO

Shared by:

Model type: Computer Vision

License:

Resources for more information:

Training Details

Training Data

The training dataset is not provided. For the evaluation dataset dataset is used.

Testing Details

Metrics

The get more information about evaluation, please check

Angle	Mean ± std. of absolute error
Yaw	5.4 ± 4.4
Pitch	5.5 ± 5.3
Roll	4.6 ± 5.6

Technical Specifications

Input/Output Details

Input:

Name: data

Info: NCHW BGR un-normalized image

Output:

Name: Multiple (please consult NN archive config.json)

Info: The three predicted values - yaw, pitch, roll.

Model Architecture

Head pose estimation network is based on simple, handmade CNN architecture. Angle regression layers are convolutions + ReLU + batch normalization + fully connected layer with one output.

Throughput

Model variant: head-pose-estimation:60x60

• Input shape: [1, 3, 60, 60] • Output shapes: [[1, 1], [1, 1], [1, 1]]

• Params (M): 1.912 • GFLOPs: 0.053

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	638.64	N/A
RVC4	FP16	670.09	2.29

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Quantization

The model is only supported in floating-point 16 and it is fast enough. It can still run on DSP.

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/head-pose-estimation:60x60"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

RegressionParser that outputs message (yaw).

RegressionParser that outputs message (pitch).

RegressionParser that outputs message (roll).

The model is multi-headed. You can set up the queues as follows:

yaw_parser_output_queue = nn.getOutput(0).createOutputQueue()
pitch_parser_output_queue = nn.getOutput(1).createOutputQueue()
roll_parser_output_queue = nn.getOutput(2).createOutputQueue()

Get parsed output(s):

while pipeline.isRuning():
    yaw_parser_output: Predictions = yaw_parser_output_queue.get()
    pitch_parser_output: Predictions = pitch_parser_output_queue.get()
    roll_parser_output: Predictions = roll_parser_output_queue.get()

Example

You can quickly run the model using script. It detects the face using model and then runs head pose detection model on the crops. To try it out, run:

python3 main.py

Lightweight CNN model predicting head pose.
License	Apache 2.0 Commercial use
Downloads	6483
Tasks	Regression
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC3, RVC4	Almost 2 years ago