3D Face Tracking and Reconstruction using Modern C++
summary
Two lightweight C++11/14 libraries implement cascaded regression landmark detection and 3D face fitting that reconstruct a face from a single 2D image, trained on hundreds to thousands of labeled images. Training matrices reach about 15,000×8,000 so MATLAB solves them in ~3 minutes while naive C++ attempts took hours, leading to Eigen and a tuned partial pivoting LU with OpenMP choice for speed. Models serialize to hundreds of megabytes using cereal serialization, the superviseddescent regressor runs live in demo, and expression handling remains a known weakness.
AI-written summary · about
chapters (10)
- 0:00Introduction and background
- 2:48Face detection and landmarks
- 8:10Linear regression for landmarks
- 16:08C++ challenges in solving equations
- 19:39Importing models from MATLAB
- 21:28Serialization with serial
- 22:533D face model reconstruction
- 24:22Live demo of tracking
- 26:07Library features and future work
- 28:56Q&A session
related talks
01h 12mBuilding a live face recognition system
52mFace Recognition: From Scratch To Hatch
15mC-Face: Continuously reconstructing facial expressions by deep learning contours of the face ...
40m"Introduction to Face Processing with Computer Vision|"
57mFoundations of Facial Recognition
39mBuilding Highly Scalable Facial Recognition Pipelines
42mDeep Learning approaches meet 3D data
21mLeveraging Python for Real-Time Image processing