Camera tracking - matchmove, in the trade - is how CG enters real footage without an audience noticing the seam. Blender's tracker has been production-grade for years; what changed recently is everything around it. Assistive trackers survive footage that used to be untrackable, segmentation tools delete moving people from plates in seconds, and cleanup passes turn grainy phone video into something a solver can love. The AI did not replace the craftsman. It took over the parts the craftsman hated.
The classic pipeline, still the spine
01 - Prepare the plate. Clean, denoised, honest-frame-rate footage. If the shot is soft and grainy, the upscale and denoise pass runs now - tracking quality is decided before the tracker opens.
02 - Track features. Set the frames, run the automatic feature detection, and then spend real minutes curating: delete trackers that wander, add manual markers on high-contrast corners, keep the survivors distributed across near and far geometry. This curation is the job - the button just starts it.
03 - Solve. Set the camera parameters - focal length if you know it, sensor size if you do not - and solve. Read the error graph like a heart monitor: flat and low is healthy, a rising curve names the frame where a tracker went bad.
04 - Ground and test. Set the floor plane and scale from a known object - a doorway, a person - then place a test cube and scrub. A cube that sits still against the plate is the entire goal of everything above.
Where AI helps, specifically
Tracking robustness. Assistive trackers hold onto features through motion blur and noise that scatter classic detection - fewer manual markers, less babysitting.Masking. Subject segmentation turns what used to be an hour of rotoscoping into a cleanup pass on an automatic matte - the single biggest time-saver in the modern stack. Plate repair. Denoise and upscale before tracking, and remove a wobble or a boom mic from the frame while you are at it. What remains firmly manual: deciding what the solve needs and judging whether it is finished.
The three failure modes and their fixes
Parallax starvation. A gentle pan across a flat wall is mathematically a rotation, not a translation - asking for a full 3D solve from it is asking for a lie. Track it as a tripod or rotation shot, or plan the scene knowing the camera is locked in place.
Rolling shutter wobble. Phone and consumer-camera footage bends straight lines as the sensor scans; the solve inherits the bend as jitter. Correct it in preprocessing, or keep the solved camera move modest enough that the wobble stays below notice.
The wrong focal length. Tell the solver you shot at 24 mm when it was 35 and every solved distance bakes in the error - the shot looks fine until the CG ghost refuses to sit still. When in doubt, let the solver estimate and compare; the error graph will pick the winner.
Set dressing and the handoff
A solved camera is an invitation: holdout masks carve the CG out of the foreground, proxy geometry receives shadows, and the shot composes. This is also where tracking meets the rest of the pipeline - a captured performer (motion capture page) into the tracked plate, an animated prop through the solved camera, a whole CG set replacing the location. The solve is the plumbing of modern VFX; now that the boring half got faster, the fun half - the part nobody trained AI to do - is all yours.