Point Track (1-4)

Planar tracking is great until there's no plane. Following a hand, the tip of a tool, a spot on a moving car - none of those are flat, textured regions you can lock onto. That's what Point Track is for: you give it one to four points in a reference frame, it follows them through the shot, and it fits the richest transform your point count allows. One point gets you translation, two get similarity (translation + rotation + scale), three get affine, four get a full homography.

How it works

Each point is tracked with Lucas-Kanade optical flow, chained forward from the reference frame. The per-frame transform is fitted directly from the reference positions to the current positions, so model error doesn't accumulate the way pure chaining would. Points the flow loses freeze at their last known position and quietly tax the confidence score - a point alive the whole shot reads 1.0; one lost halfway reads 0.5. So confidence doubles as a "did anything fall off" flag.

Inputs and outputs

• points - a JSON list of [x, y] pairs in plate pixels: [[320, 180]] for one point, [[320,180],[400,190]] for two. Multiline string, so write it however reads clearly.
• images - the frames.
• reference_frame (default 0) - the frame your points are defined on.
• downscale (default auto) - internal tracking resolution. Auto caps track width at 960px, which keeps it fast; bump to "1" for full-res if your points are small and drifting.

Outputs: track_data (which feeds Stabilize, Corner Pin, Export, the whole family) and preview, the annotated frames. Note this node doesn't carry the fancy editor widget Planar Track has - you set points as text and re-run.

Installing it

Part of the same pack. Search PlateTrack in Manager, or clone + pip install -r requirements.txt. CPU-only, no models.

Where people get burned

• Four points don't beat one if the points are bad. Points should sit on high-contrast spots (a knuckle, a logo corner, a scratch) the flow can actually see. Tracking a smooth, featureless point is like asking a tracker to follow a blank wall.
• A lost point taxes confidence but doesn't error, so a track can look "fine" while one point silently froze. Watch the confidence.
• Point tracks drift more than planar tracks on long shots - there's no ECC refine-to-reference in the point path. For a long, demanding shot, a planar track on a textured region is usually the stronger choice.

Point Track is the pragmatic fallback: 30 seconds to set up, no ROI drawing, and it slots straight into the same stabilize / corner-pin / export machinery as the heavy-duty nodes.