L2CS-Net

L2CS-Net is a gaze direction estimation model that detects faces and predicts each face's yaw and pitch angles. You can run it through our Serverless Cloud API.

L2CS-Net API

Run L2CS-Net through the HTTP endpoint directly with curl, or with the inference-sdk wrapper.

export ROBOFLOW_API_KEY="your-key-here"

curl --location 'https://serverless.roboflow.com/gaze/gaze_detection' \
  --header 'Content-Type: application/json' \
  --data '{
    "api_key": "'"$ROBOFLOW_API_KEY"'",
    "image": {"type": "url", "value": "https://media.roboflow.com/inference/man.jpg"}
  }'

export ROBOFLOW_API_KEY="your-key-here"

pip install -U inference-sdk supervision

import os
import supervision as sv
from inference_sdk import InferenceHTTPClient

image = sv.load_image_from_url("https://media.roboflow.com/inference/man.jpg")

client = InferenceHTTPClient(
    api_url="https://serverless.roboflow.com",
    api_key=os.environ["ROBOFLOW_API_KEY"],
).select_api_v1()

result = client.detect_gazes(image)

for prediction in result[0]["predictions"]:
    face = prediction["face"]
    yaw = prediction["yaw"]
    pitch = prediction["pitch"]
    print(f"Face at ({face['x']}, {face['y']}) - yaw: {yaw:.3f}, pitch: {pitch:.3f}")

L2CS-Net inference speed

Latency measured with Roboflow Inference on 1x NVIDIA L4, batch size 1, mean after warmup.

Measured on a single face crop, which is what the model expects as input.

The response contains a list with predictions (each with face bounding box, landmarks, yaw, and pitch in radians), time, time_face_det, and time_gaze_det.

For self-hosted deployments and additional examples, see the Roboflow Inference docs.

Run L2CS-Net with self-hosted Inference

L2CS-Net can also be served by a local Inference server:

pip install inference inference-cli inference-sdk
inference server start  # serves http://localhost:9001

import os
from inference_sdk import InferenceHTTPClient

client = InferenceHTTPClient(
    api_url="http://localhost:9001",
    api_key=os.environ["ROBOFLOW_API_KEY"],
)

client.detect_gazes(inference_input="./image.jpg")

The model returns one entry per detected face, containing the face box and landmarks plus yaw and pitch in radians:

[{'face': {'x': 1107.0, 'y': 1695.5, 'width': 1056.0, 'height': 1055.0,
           'confidence': 0.9356, 'class': 'face', 'class_id': 0,
           'landmarks': [{'x': 902.0, 'y': 1441.0}, {'x': 1350.0, 'y': 1449.0},
                         {'x': 1137.0, 'y': 1692.0}, {'x': 1124.0, 'y': 1915.0},
                         {'x': 625.0, 'y': 1551.0}, {'x': 1565.0, 'y': 1571.0}]},
  'pitch': 0.0295,
  'yaw': -0.0410}]

Converting yaw and pitch into a point in space assumes faces are roughly one meter from the camera and roughly 250 mm tall, which is a reasonable starting point for webcam setups.

The gaze detection example in the Inference repository shows how to run L2CS-Net on a webcam, compute where a person is looking, and annotate the frame.

Execution modes in Workflows

When used in a Workflow, gaze detection runs in one of two modes:

• Local execution: the model runs on your Inference server.
• Remote execution: the model is invoked over HTTP on a remote Inference server through the detect_gazes() client method.

Further reading

• Gaze detection and eye tracking: a how-to guide