3D Facial Landmarks Detection

A Python package for detecting and analyzing 3D facial landmarks (facemarks) from mesh data. This package provides a complete pipeline for importing 3D facial meshes, predicting facial landmark positions in 3D space, and visualizing results.

Features

• Mesh Import: Import 3D facial meshes with texture support via Open3D
• 3D Landmark Prediction: Detect 478 facial landmarks in 3D space using multi-view projection and raycasting
• JSON Export: Save predicted facemarks with normalized coordinates and closest vertex indices
• Visualization: Render meshes with detected landmarks overlaid
• Robust Detection: Uses multiple camera projections for accurate 3D reconstruction

Requirements

• Python 3.11
• MediaPipe Face Landmarker model (in working directory)

  wget https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task

Installation

Install the package via pip:

pip install facemarks

Quick Start

from facemarks import (
    import_mesh_and_setup,
    predict,
    save_facemarks_json,
    render_result
)

# Import and setup your 3D mesh
meshes = import_mesh_and_setup("path/to/mesh.obj")

# Predict 3D facial landmarks
prediction_result = predict(meshes, projections=100)

# Extract results
facemarks_3d = result["facemarks_3d"]
closest_vertex_ids = result["closest_vertex_ids"]

# or Save results directly to JSON
save_facemarks_json(
    "path/to/mesh.obj",
    prediction_result,
    "output/facemarks.json"
)

# Visualize the results
render_result(meshes["textured"], facemarks_3d)

API Reference

import_mesh_and_setup(filename)

Imports a 3D facial mesh and prepares it for landmark detection. Supports textured OBJ files.

Parameters:

• filename (str): Path to the mesh file (.obj format)

Returns:

• dict: Dictionary containing mesh data with keys:
  • "original": The base mesh geometry
  • "textured": Textured mesh for rendering
  • "tensor": Tensor representation for raycasting

Example:

meshes = import_mesh_and_setup("face_model.obj")

---

predict(meshes, projections=100)

Predicts 3D facial landmarks from the provided mesh data using multi-view projection and raycasting. Detects 478 facial landmarks in 3D space.

Parameters:

• meshes (dict): Mesh data object returned from import_mesh_and_setup()
• projections (int, optional): Number of camera projections to use for landmark detection. Default is 100. More projections increase accuracy but take longer.

Returns:

• dict: Dictionary containing:
  • "facemarks_3d" (list): List of 3D coordinates for detected facial landmarks
  • "closest_vertex_ids" (list): Indices of the closest mesh vertices to each landmark

Example:

result = predict(meshes, projections=150)
landmarks = result["facemarks_3d"]
vertex_ids = result["closest_vertex_ids"]

Note: The function uses random camera rotations for robust multi-view detection. If fewer than half of the requested projections successfully detect a face, an error message is printed.

---

save_facemarks_json(input_path, prediction_result, json_path)

Exports predicted facial landmarks to a JSON file with normalized coordinates and vertex mapping.

Parameters:

• input_path (str): Path to the original input mesh file
• prediction_result (dict): Prediction data object returned from predict()
• json_path (str): Destination path for the output JSON file

Output JSON Structure:

{
    "model": "path/to/mesh.obj",
    "normalized coordinates": [[x, y, z], ...],
    "closest vertex indexes": [idx1, idx2, ...]
}

Example:

save_facemarks_json(
    "input/face.obj",
    prediction_result,
    "output/landmarks.json"
)

---

render_result(mesh, facemarks)

Visualizes the mesh with predicted facial landmarks overlaid as magenta points.

Parameters:

• mesh: The mesh geometry to be displayed (either from meshes["original"] or meshes["textured"])
• facemarks (list): List of 3D landmark coordinates to visualize

Example:

render_result(meshes["original"], result["facemarks_3d"])

Note: This function requires a display environment. It will not work in headless environments with virtual displays.

How It Works

The package uses a sophisticated multi-view approach to detect 3D facial landmarks:

1. Multi-View Projection: The mesh is rendered from multiple random camera angles (default: 100 views)
2. 2D Landmark Detection: MediaPipe Face Landmark detection is applied to each rendered view
3. Ray Casting: Rays are cast from camera positions through detected 2D landmarks onto the 3D mesh
4. 3D Reconstruction: The intersection points are aggregated across all views to compute robust 3D landmark positions
5. Vertex Mapping: Each landmark is mapped to the closest vertex on the original mesh

Use Cases

• Facial animation and rigging
• 3D face model analysis
• Biometric applications
• Character modeling pipelines
• Medical and dental visualization
• Motion capture alignment

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Support

For issues, questions, or contributions, please visit the GitHub repository.