Proposal: Hybrid Kalman Tracking for 1.2x Speedup on CPU

[safhac]

Hi py-feat team,
I have been working on optimizing py-feat for video processing on CPU/low-resource environments.
Currently, the detector runs heavy models (like img2pose or RetinaFace) on every single frame, which is computationally expensive even when face positions are predictable.

The Solution I implemented a Hybrid Tracking Engine that decouples "Search" (Detection) from "Tracking."

1. Kalman Filter Integration: Uses filterpy to track face coordinates across frames.

2. Hybrid Switching: Added a detection_interval argument.
   The heavy detector only runs every $N$ frames (e.g., 5) to correct the tracker. Intermediate frames are predicted via Kalman Filter (CPU).

3. Adaptive Padding: Implemented a safety crop (20% padding) to prevent the "drifting crop" issue where landmarks fail if the tracker is slightly misaligned.

Benchmark ResultsTested on a simulated workload (0.1s load per detection) over 50 frames:

1. Running BASELINE (Interval=1)...
   -> Time: 31.91s | FPS: 1.57

2. Running HYBRID (Interval=5)...
   -> Time: 27.63s | FPS: 1.81

SPEEDUP ACHIEVED: 13.4%
PERFORMANCE BOOST: 1.2x Faster

Note: The speedup affects the Detection stage. The Analysis stage (Landmarks/AUs) still runs on every frame to ensure accuracy.

Code You can view the implementation in my fork here: Link to feat/detector.py]

Next Steps I have this working and tested on my fork. Would you be open to a PR for this feature?