Planar Tracking

How to use it

1. 1 Open MotionMaster 3D Planar & Tripod Trackers, choose Planar, and pick an engine: Production CUDA (NVIDIA, default), TAPNext++ Hybrid, or Classic CPU.
2. 2 Choose the motion model (Translation, Similarity, Shear / Scale, or Perspective), the quality profile, the frame range, and the reference frame, then click Open Track Workspace.
3. 3 Draw the planar quad over the surface on the reference frame. Corner order is Upper Left, Upper Right, Lower Right, Lower Left.
4. 4 Track forward, backward, or both from the current frame. Every accepted frame is committed immediately, so a later failure keeps everything solved before it.
5. 5 If a corner slips, set a protected manual key or place AdjustTrack reference points and move them to the true position; the raw solve is never overwritten.
6. 6 Choose how missing or off-screen intervals are bridged: Interpolate and Hold Ends, Interpolate and Extrapolate, Hold Nearest, or Require Measured Frames.
7. 7 Publish to Blender as Object Motion (camera-parented animated plane) or Camera Motion (fixed plane plus animated camera). Add further surfaces into the same camera with Add To Existing Camera.

Technical notes

Engines: Production CUDA is a native C++/CUDA fixed-reference subpixel plane tracker with round-trip point consensus, global CUDA SIFT re-detection, and hard anti-drift acceptance gates. TAPNext++ Hybrid adds TAPNext++ long-track guidance and XFeat/LighterGlue re-detection on top of the same fixed-reference refinement; it needs CUDA PyTorch and a one-time verified model download of about 2.4 GB from the panel. Classic CPU uses OpenCV pyramidal tracking with MAGSAC++, SIFT, and ECC and runs on any machine. Pixel-Accurate Plane Lock settings (motion model, input channel, and percentage of strongest pixels used) apply to the CUDA and Hybrid engines.

Every frame is measured for blur, directional motion blur, grain, and contrast. Fast motion uses a bounded motion prediction to seed optical flow, forward/backward checks and MAGSAC decide which correspondences are credible, and low-contrast or grain-heavy footage receives scheduled reference-frame validation so drift is corrected before it accumulates. A large disagreement between incremental and reference motion requires independent confirmation; if neither hypothesis can be proven, tracking stops with measured diagnostics and saves every accepted frame instead of publishing drift.

Corners may leave the image. Fully off-screen surfaces cannot be measured, so generated bridge frames are stored with low-confidence review diagnostics and are never labelled as measured tracks. Publish Result fills gaps automatically with the selected policy, and editing a generated frame promotes it to a protected manual key.

Publishing always solves through MotionMaster 3D's own camera and never adopts a camera you placed. Estimate Focal Length is on by default and is measured from genuinely tracked frames only; turn it off to apply and lock a manual focal length in millimetres. A Camera Motion publish leaves exactly one camera in the scene. Camera Motion uses IPPE planar pose candidates with nonlinear refinement and temporal continuity, and Plane Depth scales the plane and camera together so framing stays stable.

Outputs are one visible native four-corner Plane Track, an animated compositor Corner Pin node with a connected Movie Clip source, and a four-vertex 3D plane. A portable planar_track.json (schema motionmaster3d.planar_track.v1) is written to a versioned planar_track_result folder, so Blender outputs can be recreated from Import Results without retracking.

Planar tracking assumes the selected pixels belong to one physical plane. Folded, deforming, heavily reflective, textureless, or fully occluded surfaces stop the automatic tracker at its quality gate; bridge those intervals with manual keys and an explicit gap policy.