Gaze estimation ADAS

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

1+ Likes

Model Details

Model Description

Gaze estimation model is a lightweight, simple, and handmade convolutional neural network. The model has 3 inputs: cropped left-eye image, cropped right-eye image, and 3 values (pitch, yaw, roll) from head pose estimation model. The network outputs 3-D vector corresponding to the direction of a person's gaze in a Cartesian coordinate system in which Z-axis is directed from person's eyes (mid-point between left and right eyes' centers) to the camera center, Y-axis is vertical, and X-axis is orthogonal to both z,y axes so that (x,y,z) constitute a right-handed coordinate system.

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 two random held out inviduals from an OpenVINO internal dataset containing images of 60 people with different gaze directions are used.

Testing Details

Metrics

The accuracy of gaze direction prediction is evaluated through the use of Mean Absolute Error of angle (in degrees) between the ground truth and predicted gaze direction.

Angle	Mean ± std. of absolute error
OV Internal Dataset	6.95 ± 3.58

Technical Specifications

Input/Output Details

Input:
- Name: Multiple (please consult NN archive config.json)
Output:
- Name: Identity
  - Info: 3D coordinates of gaze estimation.

Model Architecture

The model is a simple, custom VGG-like convolutional neural network with 3 inputs and 1 output.

Throughput

Model variant: gaze-estimation-adas:60x60

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

• Params (M): 1.882 • GFLOPs: 0.070

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	564.65	N/A
RVC4	FP16	656.25	2.45

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

* Parameters and FLOPs are obtained from the package.

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

This model should work together with face detection and head pose estimation models.

Define model:

model_description = dai.NNModelDescription(
    "luxonis/gaze-estimation-adas:60x60"
)

The model expects 3 inputs (multi-input model): cropped left eye image, cropped right eye image and head pose values. You can check the example on how to use .

Inspect model head(s):

RegressionParser that outputs message (3D vector).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: Predictions = parser_output_queue.get()

Example

You can quickly run the model using our example.

The example demonstrates how to build a 3-stage DepthAI pipeline consisting of a face detectoion model, head pose estimation model, and a gaze estimation model. It automatically downloads the models, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool.

To try it out, run:

python3 main.py

Fast Gaze estimation model.
License	Apache 2.0 Commercial use
Downloads	970
Tasks	Regression
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC4	8 months ago