Tool 26 · Session 3 · Video · Point correspondence
Point Correspondence Lab
Select a visible point in frame one, predict where it will be in frame eight, then reveal a precomputed track. Compare a clean path with occlusion, identity-switch, and off-screen failures.
Big idea: PoseNet can estimate a keypoint in one frame; tracking asks which later point corresponds to it. Generative video has the harder inverse problem—create each new frame while keeping those relationships coherent.
Authored teaching simulation. The scenes, routes, confidence values, and failures are precomputed illustrations. They are not PoseNet or CoTracker outputs, and they do not measure a live model.
Clean correspondence
Follow the point, not the object name
simulated tracker estimate
expected identity / location
Choose one point in frame 1.
Frame 1 of 8
Illustrated confidence
96%
Investigation note
Make one bounded claim
Describe what this authored route illustrates. A confidence value is part of the simulation—not proof that the point is correct.
Mechanism boundary
Tracking is not generation
| Pose estimation | Estimate named body keypoints in one existing frame. |
|---|---|
| Point tracking | Estimate where a selected point went across frames that already exist. |
| Video generation | Create later frames while making identities, locations, motion, camera, and physics remain believable. |
| Human decisions | Which points matter, what counts as the same feature, which failures are tolerated, and how confidence is interpreted. |
Debrief questions
- What evidence made you call two points “the same” across time?
- Did confidence fall before, during, or after the visible failure?
- Why can a plausible point still be the wrong identity?
- What additional relationships must a generator preserve beyond this one point?