Camera Tracker Node

The Camera Tracker is a structure-from-motion solver. It analyzes how features move through the footage, then solves all of these at once:

• The 3D path of the camera (translation + rotation on every frame).
• A 3D point cloud of the scene - tracked points rebuilt in space.
• Optionally, the camera's focal length if it is unknown.

With this exported data, you can place 3D elements (models, text, particles) that really look like they were filmed in the scene - they stay firmly anchored to the ground or to real objects while the camera moves.

• Effects Library: Tools > Tracking > Camera Tracker
• Quick add: right-click in the Node Editor → Add Tool > Tracking > Camera Tracker

Classic Tracker

For tracking one or a few 2D points. Stays in 2D.

Planar Tracker

For flat surfaces (screen, sign, road). Stays in 2D but with perspective.

Camera Tracker

For solving the camera in 3D. It leaves 2D behind - result: a 3D scene with an animated camera + a point cloud. For integrating CGI and anything projected in 3D.

Rule of thumb: if the end goal is an element that moves with the scene (CG, volumetrics, 3D matte painting), use the Camera Tracker. If it is 2D screen replacement, use Planar.

• Background (orange) - the footage to track.
• Occlusion Mask (white) - mask of the areas to ignore: moving people, vehicles, water, clouds. Crucial, because their points create false tracks that ruin the solve.

MediaIn → Camera Tracker → Export
                       ↓
                 Camera 3D, Point Cloud 3D, Shape 3D (ground),
                 Merge 3D, Renderer 3D, Merge 2D → MediaOut

With occlusion (a person):
MediaIn → Camera Tracker ← Occlusion Mask (roto of the person)

1. Track Phase

   Automatic detection of trackable features (through pattern matching). Adjust the sensitivity and the feature spacing. Run Bidirectional Tracking to get longer tracks (backward after forward).

2. Solve Phase

   The solver rebuilds the camera + point cloud from the tracks. Provide the focal length if you know it (EXIF, camera data) for a better result. Otherwise Fusion tries to compute it - it works but is less precise.

3. Refine Phase (the critical step)

   An iterative cycle:

   Track → Solve → Refine Filters → Solve →
   Cleanup tracks → Solve → Cleanup from point cloud → Solve → Repeat.

   Filter the tracks by minimum length, tracking error and solve error. Delete the tracks coming from moving objects, reflections and parallax errors. This is the most important step - a clean solve around 0.6-0.8px only comes from these cleanup cycles.

4. Export Phase

   Automatically generates: an animated Camera 3D, Point Cloud 3D, Shape 3D (ground plane), Renderer 3D, Merge 2D - a complete chain ready for CG integration.

< 0.5 px = excellent

Near-perfect tracking. Usable as is.

0.6 - 0.8 px = very good

Pro quality, publishable result. The target to aim for.

0.8 - 1.0 px = good

Acceptable for most needs. Slightly visible if you zoom in.

> 1.0 px = problem

Visible drift in the render. Keep refining: delete short tracks, outliers and moving objects.

Clicking Export creates a ready-to-use node structure:

• Camera 3D - a virtual Fusion camera with animated position/rotation matched to the real camera.
• Point Cloud 3D - the cloud of tracked points in the rebuilt 3D space. Used to line up CG objects.
• Shape 3D (ground plane) - a default plane placed at the estimated ground level.
• Merge 3D - the scene that assembles all the 3D elements.
• Renderer 3D - renders the 3D scene to 2D, aligned with the original shot.

• Integrating a CG object - a robot, car or 3D creature placed in the live scene, staying anchored in place while the camera moves.
• 3D matte painting - a 3D digital backdrop projected through the tracked camera, consistent with the perspective and the motion.
• Volumetric effects - particles, smoke, fire, light rays that really exist in 3D in the scene and do not slide like a 2D overlay.
• Laser beams, bullets, projectiles - 3D paths that cross the scene with the right perspective and anchoring.
• Hologram / 3D UI - floating screens with text or graphics that respect the depth of the scene.
• Set extension - extend a real set with 3D elements (walls, buildings, horizon) that hold up in perspective.
• 3D sky / ground replacement - sky or ground replaced with 3D elements (not just a 2D screen replacement but a real 3D plane with parallax).

• Long tracks = a better solve - prefer fewer but longer tracks (over 50+ frames) to lots of short ones.
• Distant seed frames - two reference frames well apart in the shot (not neighbors) + that share as many tracks as possible. This enables precise triangulation.
• Focal length provided = a precise solve - get the focal length of the take (EXIF, camera report) and enter it. A significant gain in quality.
• Occlusion mask required for people - always roto moving objects and connect them to the Occlusion Mask. Otherwise the points on them create false tracks that make the solve drift.
• Do not over-track - too many tracks = noise. Aim for a few hundred at most, not thousands.
• Shots that are hard to track:
  • Locked-off shots (no parallax → no solve possible).
  • Pure zoom moves (no translation).
  • Footage full of motion blur.
  • Textureless surfaces (sky, snow).
  • Reflections (water, glass, mirror).
• Keep refining until it plateaus - the first solve is rarely good. Each cleanup cycle brings the error down by a few tenths. Stop when the gain becomes marginal.
• For difficult shots (pro VFX), prefer dedicated solutions (SynthEyes, 3DEqualizer, Mocha Pro) that offer more control. Fusion's Camera Tracker is excellent for 80% of cases and built in.