A 2009 face-tracking paper, still compiling in 2024

What it does

FaceTracker maps facial landmarks in real time from a webcam feed. It bundles a pre-trained deformable model, a reference CLI tool, and enough wrappers (openFrameworks, Cinder, Python, Android) to suggest it once powered a small ecosystem of art installations and prototypes.

The interesting bit

The core algorithm—“Face Alignment through Subspace Constrained Mean-Shifts”—dates to ICCV 2009. That makes this a fossil record of how computer vision worked before deep learning ate the field: hand-crafted models, iterative optimization, and actual C++ you can step through without a GPU.

Key highlights

• MIT-licensed since 2020; previously paywalled for commercial use
• Ships with a pre-trained model (face2.tracker) and triangulation data
• Optional OpenMP support for parallelization
• Wrappers exist for mobile, creative-coding frameworks, and Python
• CLI includes a rescaling flag (-s .25) because 720p webcams will swamp the detector

Caveats

• Explicitly unmaintained: README states OpenCV 4 broke the build and the maintainers gave up
• Build instructions reference homebrew/science, a tap Homebrew retired years ago
• The FAQ practically apologizes for crashes, slow detection, and high CPU usage

Verdict

Worth a look if you’re researching pre-neural face tracking, maintaining legacy interactive art, or want a readable C++ baseline to compare against modern mediapipe. Everyone else should probably just use mediapipe.