3D Face Reconstruction From Stereo Images

This project focuses on 3D face reconstruction from stereo images under three calibration scenarios: fully calibrated, partially calibrated, and uncalibrated.

A calibrated 3D face was set up in Blender with three cameras capturing the face from left, front, and right angles. The front camera remained fixed as a reference, while the side cameras varied in pose across four configurations to evaluate the effects of rotation and translation on reconstruction.

Calibration Setups

The datasets were captured under four calibration setups, which are as follows:

• Dataset 0: Translation of ±0.1 m without rotation
• Dataset 5: Rotation of ±5° combined with ±0.1 m translation
• Dataset 10: Rotation of ±10° combined with ±0.2 m translation
• Dataset 20: Rotation of ±20° combined with ±0.4 m translation

Each dataset contains three images, three ground truth depth maps, and a JSON file with all camera parameters.

Four separate programs are provided for calibrated, uncalibrated, and partially calibrated (pose estimation) reconstruction, as well as depth map visualization.

Each program allows you to change the dataset and observe how camera pose affects the accuracy of 3D reconstruction.

Installation

• git clone https://github.com/kidat/3D-Face-Reconstruction-from-Stereo-Images.git 3DFace
• cd 3DFace
• pip install -r requirements.txt

🔽 Downloadable 3D Face Result Examples

✅ Calibrated

• Calibrated_reconstruction.ply
• Calibrated_reconstruction_textured.obj

🌐 Depth Map

• DepthMap.ply

📍 Partial Calibration (Pose Estimation)

• Reconstruction_pose.ply
• textured_mesh_pose.ply

🚫 Uncalibrated

• reconstruction.ply

💡 You can view these files using MeshLab, CloudCompare, or any 3D viewer that supports .ply or .obj.

License

MIT License - See LICENSE file for details