Locked to the Ground

How 3D text, objects and whole CG layers are placed into real footage so they stay put while the camera moves. How camera tracking rebuilds a camera from nothing but the picture, what sells the final composite, and the software that does it, from free beginner tools to the Hollywood standard.

Post-production · VFX · 19 min

Every week you watch things that were never filmed. A title that stands on a street in the opening credits, a logo sitting on a table in an advert, a spaceship parked behind a real actor, a whole city added behind a real rooftop. In each case a 3D layer made in a computer has been placed into live footage so well that it seems to have been in front of the lens. The hard part is not making the 3D object. The hard part is making it move exactly the way the real camera moved, frame after frame, so it never slides. This post explains how that is done: the problem, how camera tracking (also called matchmoving) rebuilds the real camera from the picture, how artists light, shadow and grade the result until it belongs, and which software you can do it in, from free tools for beginners to the programs used on blockbuster films. Every figure is drawn live in your browser from one small 3D scene, so you can scrub through the shot and watch the tracker work.

Why a pasted layer slides

Compositing means stacking layers, like sheets of glass. The bottom layer is the **plate**, the real footage. On top go the extra elements: text, a 3D model, smoke, a matte painting. If the camera was locked off on a tripod, the plate never moves and you can simply paste a picture of your 3D object on top. The moment the camera moves, this breaks. When a camera moves, nothing in the picture moves by the same amount. Things close to the lens rush past, things far away barely move, and the floor stretches and turns. A pasted layer has none of that. It hangs on the glass in front of the scene, and your eye notices at once that it is sliding over the ground instead of standing on it.

The fix is to film the 3D object with a camera that moves exactly like the real one. If the virtual camera in your 3D software has the same position, the same rotation and the same lens as the real camera on every single frame, then a 3D object rendered through it will shift, shrink and turn in perfect agreement with the plate, and it will look glued to the ground. Finding that virtual camera is the job called **camera tracking** or **matchmoving**. Productions sometimes record camera data on set with motion-control rigs or encoders, but most of the time the only record of how the camera moved is the footage itself. Matchmoving software reads that footage and works the camera out backwards.

How a camera is rebuilt from a picture

Step 1 · Follow the dots

Tracking starts in 2D. The software looks for small patches of the image that are easy to recognise, usually corners where light meets dark, such as the end of a road marking, the corner of a window or a crack in the pavement. Each patch is called a **feature**. A small box around it is the **pattern**, and a larger box around that is the **search area**. On the next frame the tracker slides the pattern across every position in the search area and scores how similar the pixels are. The best score is where the feature went. Repeat that for every frame and you get a **track**: a line of 2D positions describing how one point on the set travelled across the screen. The figure above runs exactly that search on the rendered frames, live.

Not every patch is a good feature. A sharp corner can only match in one place, so it locks. A straight edge matches anywhere along its length, so the tracker slides along it; this is called the aperture problem. Plain sky or a white wall has nothing to hold on to. Repeating patterns such as fences, tiles and brick can make the tracker jump to a neighbour that looks the same. And some features track perfectly but move the wrong way: people, cars, trees in the wind, reflections and highlights that slide over shiny surfaces as the camera moves. These do not belong to the still set, so they would confuse the camera solve. Artists draw **masks** over moving things so the tracker ignores them, and they delete bad tracks before solving.

Nobody places hundreds of trackers by hand any more. Every modern camera tracker has an **automatic** mode that scatters features over the whole frame, tracks them all, and lets you remove the bad ones. Artists then add **supervised** tracks by hand where they need extra precision, for example exactly where a 3D object will touch the ground. Tracks are born when a feature enters the frame and die when it leaves or gets hidden, so the count changes all the time. What the solver needs is enough tracks on every frame, spread across the picture and at different distances from the camera. As a rough guide, the Blender manual says a camera solve needs at least eight tracks in common between its two keyframes. In practice artists aim for far more than that so the solve has spares when some tracks go wrong.

Step 2 · Parallax tells depth

A single 2D track says nothing about distance. The point could be a pebble near the lens or a mountain on the horizon. Depth comes from **parallax**: when the camera physically moves, near things move across the screen faster than far things. Look out of a train window and the fence posts race by while the hills crawl. This is the most important idea in camera tracking. A shot where the camera travels, such as a dolly, a crane, a drone or a walking handheld shot, has parallax and can be solved into a full 3D camera with depth. A shot where the camera only pans and tilts on a tripod has almost none. The solver can still recover the rotation and the lens, but not the distances, so tools offer a special **tripod** or nodal solve for those shots.

Step 3 · Solve the camera

Now the clever part. The software has hundreds of tracks and knows nothing else. It guesses a camera position and rotation for every frame and a 3D position for every tracked point. Then it projects each 3D point through each guessed camera and checks where it lands on the screen. Where the projected point misses its tracked position, the guess is wrong, so the solver nudges the cameras and the points to reduce the misses and tries again, thousands of times. The method is called **structure from motion**, and the final refinement step is called **bundle adjustment**, because it adjusts the bundles of light rays from every point to every camera at the same time. When it converges you have two results: an animated camera with a position, rotation and focal length for every frame, and a **point cloud**, a sparse 3D copy of the set made of the tracked points.

The solver reports how well it did as a **solve error**, also called reprojection error. It is the average distance, in pixels, between where each point was tracked and where the solved camera says it should be. Adobe's After Effects calls it "Average Error", and Nuke reports it as an RMS value. Foundry's Nuke documentation suggests fine-tuning a solve whose error is above one pixel, and guides for Fusion's Camera Tracker give the same rule of thumb for HD footage. Lower is better, but a low number is not the whole story. A solve can score well and still have the wrong scale or a tilted floor, so artists always test it by eye, placing simple test objects such as cones or a checkerboard grid on the ground and watching whether they stick for the whole shot. When the error is high, the usual cause is a group of bad tracks, and deleting them and solving again fixes most problems.

Step 4 · Know your lens

The solver assumes a perfect pinhole camera in which straight lines stay straight. Real lenses are not perfect. Wide lenses usually bow straight lines outwards near the edges (barrel distortion) and some longer lenses pinch them inwards (pincushion). Anamorphic lenses add their own squeeze and curves. If the distortion is ignored, tracks near the edges disagree with tracks in the middle and the solve fails or wobbles. So matchmovers measure the lens, ideally by filming a grid chart with the same lens on set, or let the solver estimate it from the shot. The usual workflow is to **undistort** the plate, track and render the 3D layer in that clean, straight world, then **redistort** the render so it bends exactly like the original footage before it is composited. Nuke, for example, can pass this warp between programs as an image called an ST map.

The solver also needs to know the **focal length** and the size of the camera's **sensor** (called the filmback). These two numbers decide the field of view. A wide lens close to an object and a long lens far away can frame that object at exactly the same size; only the background and the parallax tell them apart. Most trackers can estimate the focal length, and some can follow a zoom, but every number you give them makes the solve more reliable. That is why matchmove and VFX teams keep camera reports on set: the camera body, the lens and focal length for every take, the height of the camera, and measurements of the set. A tape-measure distance between two marks on the floor is often what gives the finished solve its true scale.

Step 5 · Ground, origin and scale

A fresh solve has the right shape but no idea which way is up. The camera path and the point cloud can float at any tilt, anywhere, at any size, because the footage alone cannot tell a model of a plaza from a full-size plaza. So the artist orients the scene. They pick three or more tracked points that lie on the floor and tell the software they define the **ground plane**, choose one point to be the **origin** (0, 0, 0), and set the **scale** from a known distance. After Effects does this with a right-click option called "Set Ground Plane and Origin". In Blender, Nuke, SynthEyes and the others the steps have different names but do the same job. Get this right and a 3D object placed at the origin sits flat on the real floor at the right size, and gravity in a simulation points the right way.

Other kinds of track

A camera track recovers the camera from the still parts of a set, but it is not the only kind. An **object track** solves the movement of something that moves on its own, such as a car you want to add a 3D bumper to or a box an actor carries. A **planar track** follows a flat surface, like a phone screen, a poster or a wall, and calculates how it bends in perspective so you can pin a new image onto it; it needs no 3D solve at all. Tracking faces and bodies is its own speciality. Many shots use several kinds together: a camera solve for the set, an object solve for the moving car and planar tracks to clean up a sign.

Putting the 3D layer in

Once the camera is solved and oriented, it is exported to wherever the 3D work happens. In After Effects and Blender the camera is already inside the same program. In bigger pipelines the tracker exports the camera and point cloud as an FBX, Alembic or USD file, or as a script for Maya, Houdini, Cinema 4D or Nuke. The 3D artist builds or imports the object, puts it where it belongs using the point cloud as a guide, lights it and renders it through the solved camera. The render is usually split into **passes** (the object itself, its shadow, reflections, a depth pass) so the compositor can adjust each one separately. The compositor then layers those passes over the undistorted plate, redistorts them, and finishes the shot.

With a good solve, placing the object is the easy and satisfying part. You drop it on the ground grid and it stays there through the whole shot. For **3D text** the letters are extruded to give them depth, often with a small bevel on the edges so they catch the light. In After Effects you can right-click a target in the 3D Camera Tracker and choose "Create Text and Camera", and the text appears on the plane you picked; Cinema 4D and Blender can extrude text into full 3D geometry. For a **3D object** the point cloud shows where the real surfaces are, so a robot's feet can land on the real pavement and a logo can stand exactly on the edge of a real table. The object must be built at the scale you set when orienting the solve. A one-metre crate that is accidentally ten metres tall is the most common reason a good track still looks wrong.

A perfectly tracked object can still look fake if it is lit differently from the plate. The brain is very good at noticing shadows that point the wrong way. So VFX crews record the light on set. They photograph a **grey ball**, which shows the direction, softness and colour of the light, and a **chrome ball**, which reflects the whole surroundings. They also shoot an **HDRI**, a 360-degree panorama with a very wide range of brightness, which the 3D software can use to light the object with the real environment. The artist then matches the main light's direction to the real shadows, its softness to the real shadow edges and its colour to the real highlights. With only the footage to go on, the shadows already in the plate are the best guide.

Shadows are what make an object feel heavy. But the real ground is footage, not 3D geometry, so how does a CG object cast a shadow on it? The answer is a **shadow catcher**: an invisible 3D plane placed exactly on the solved ground, which renders only the shadows that fall on it. Composited over the plate, it darkens the real paving under the object. After Effects can create one from the 3D Camera Tracker in one step, and Blender's Cycles renderer has a shadow catcher setting for any object. The finishing touch is a **contact shadow**, a soft dark line where the object meets the floor, usually from an ambient occlusion pass. Without it, even a well-tracked object seems to hover a millimetre above the ground.

A composite is a stack, and the top layer always wins. If anything real passes in front of the 3D object, such as a bollard, a lamppost or a passing actor, the 3D layer will wrongly cover it. The fix is to cut the foreground thing out of the plate with a **matte** and lay it back on top. Artists draw these by hand, frame by frame, in a process called **rotoscoping**, or they use keyers and newer AI-assisted tools that separate the subject from the background. A cleaner trick for solid objects is a **holdout**: model a simple stand-in for the bollard in 3D, place it on the solved point cloud, and render it as an invisible cut-out that hides the parts of the 3D object behind it. Either way, the point cloud from the solve tells you what is in front of what.

The last few percent of realism come from making the 3D layer as imperfect as the footage. Real cameras have **grain** or sensor noise, so the compositor adds matching grain to the CG. Real lenses are slightly soft, so perfect CG edges get a touch of blur, and objects out of the focus plane get **defocus**. A moving camera smears everything a little, so the CG needs the same **motion blur**. The colour has to be graded to match: the darkest black in the CG should be no deeper than the darkest black in the plate, and the highlights should have the same colour as the real ones. Then the lens distortion goes back on, and sometimes a **light wrap**, a glow of the background bleeding around the CG edges. Each step is small. Together they are the difference between "nice render" and "I can't tell".

From first track to feature film

Almost every program that handles video and 3D now has a camera tracker of some kind, and they share the same ideas you just saw: track, solve, orient, export. What changes as you climb is control. Beginner tools do almost everything automatically and work well on good footage. Professional tools give you full control over every track, lens model and constraint, so they can rescue hard shots such as fast handheld moves, heavy motion blur, anamorphic lenses, or a set with almost nothing to track. Here are nine common tools, roughly from easiest to most specialised. Feature descriptions are what each maker says about its own product as of October 2026; check the maker's site for current versions and pricing.

Beginner · AdobeAfter Effects · 3D Camera Tracker

For most motion designers the first camera track happens in After Effects. Adobe describes the **3D Camera Tracker** effect as analysing video to extract camera motion and 3D scene data. You apply it to the footage, it analyses in the background, and the solved points appear as small coloured crosses. You select three or more points, a target appears on the plane they define, and a right-click creates a text layer, solid, null or shadow catcher with a matching 3D camera. Adobe's documentation lists an "Average Error" readout in pixels and shot types for a fixed lens, a zoom, or a known angle of view, plus the "Set Ground Plane and Origin" command. It is fast and needs no 3D knowledge, which makes it ideal for 3D titles and simple objects; it offers less control when a shot is difficult.

Beginner · free · Blender FoundationBlender · Motion Tracking

**Blender** is free and open source, and it is the best place to learn real matchmoving because it shows you every step. Its Movie Clip Editor has a full motion tracking workspace: you place markers or detect features automatically, track them, set the camera's sensor and focal length, and choose a lens distortion model. The manual describes solving camera motion, tripod solves for pans, and object motion solving. One button, "Setup Tracking Scene", builds a 3D scene with the solved camera, the footage as a background and a ground plane that catches shadows. Because Blender is also a complete 3D modeller and renderer with a compositor, a beginner can track, model, light, render and composite a whole shot without leaving the program.

Intermediate · Blackmagic DesignDaVinci Resolve · Fusion Camera Tracker

DaVinci Resolve combines editing, colour, audio and VFX in one program, and its **Fusion** page is a node-based compositor with a real 3D workspace. Fusion's **Camera Tracker** node works through tabs in order: track the features, describe the camera and lens, solve, then export a 3D scene with the camera, point cloud and ground plane as connected nodes. The free version of Resolve includes Fusion with point and planar tracking, but the Camera Tracker node requires the paid DaVinci Resolve Studio (or the standalone Fusion Studio). It is a strong middle step: more control than After Effects, and a natural home for anyone who already edits and grades in Resolve.

Intermediate · MaxonCinema 4D · Motion Tracker

**Cinema 4D** is a favourite 3D program among motion designers, and its **Motion Tracker** object brings camera tracking into the same scene where the 3D work happens. Maxon's documentation describes automatic and manual 2D tracking, a 3D camera solve, and constraint tags that define a floor plane, an origin, a vector and a known distance, which is exactly the orienting step from Figure 12. It also has an Object Tracker for moving objects and a Camera Calibrator for working from a single still image. Because Cinema 4D connects closely with After Effects, a common motion design workflow is to track in either one and build polished 3D text and logos in Cinema 4D.

Professional · Boris FXMocha Pro

**Mocha Pro** from Boris FX is built around **planar tracking**. Instead of following tiny points, you draw a shape around a flat area, such as a wall, a screen or a tabletop, and Mocha tracks the whole surface, which makes it robust on blurry or partly hidden shots where point trackers struggle. Boris FX calls it Academy Award-winning planar tracking and lists rotoscoping, object removal, stabilisation, PowerMesh tracking for warping surfaces like skin and cloth, and 3D camera solves "powered by SynthEyes". It runs on its own or as a plugin inside hosts such as After Effects, Premiere Pro, Nuke and Flame, which is why it shows up at every level, from freelancers to big facilities.

Hollywood standard · FoundryNuke · CameraTracker

**Nuke** from Foundry is the compositing program most associated with high-end film and television VFX, and its **CameraTracker** node is available in NukeX and Nuke Studio. Foundry's documentation describes three steps: Track, which auto-tracks features; Solve, which calculates the camera and a 3D point for each track; and Create Scene, which builds an animated camera, point cloud and cards. It can estimate an unknown lens's distortion during the solve, accepts a mask to keep moving objects out, and reports an RMS error, with fine-tuning suggested above one pixel. Because Nuke's own 3D system can render geometry through the solved camera, compositors often use it for set extensions, clean-up and projections without sending the shot to a separate 3D department.

Professional · Boris FXSynthEyes

**SynthEyes** is a standalone matchmoving application, originally from Andersson Technologies and now sold by Boris FX. Boris FX describes it as production-grade 3D tracking and camera solving, with automatic and supervised tracking, multi-object and stereoscopic tracking, 360° VR, lens calibration for anamorphic and fisheye lenses, "geometric hierarchy" tracking for jointed objects, and exporters for Blender, Houdini, Nuke, After Effects, Maya, Cinema 4D and USD. It has long been known as a powerful tracker at a price freelancers and small studios can afford, which makes it the usual step up for an artist whose shots have outgrown the built-in trackers.

Hollywood standard · Science.D.Visions3DEqualizer

**3DEqualizer** from the German company Science.D.Visions is the tool many large VFX studios use in their matchmove departments. The maker calls it "the worlds leading 3D tracking solution" and says it has been the main 3D tracking tool on more than 18 Academy Award-winning films since 1998. Its pages stress lens distortion handling for spherical, anamorphic and fisheye lenses, a solver that weights every point frame by frame, and the use of lidar scans and survey data of the set. Matchmove artists in big facilities spend whole days on single shots in tools like this, tracking by hand where needed, because a feature film can show a shot on a cinema screen where half a pixel of slip is visible.

Professional · The Pixel FarmPFTrack

**PFTrack** from The Pixel Farm is a node-based matchmove and reconstruction application. Every step, such as importing the clip, auto tracking, user tracking, solving the camera, orienting the scene and exporting, is a node in a tree, so you can try different approaches side by side. The Pixel Farm describes camera tracking with calibration and survey tools, lens distortion management, photogrammetry, lidar integration and scene reconstruction, including a "Hero Cloud" node that builds dense point clouds from a single plate. It comes in Solo, Studio and Enterprise editions and is used in VFX alongside forensic, architecture and heritage work.

Side by side, the pattern is clear. The free and built-in tools, After Effects and Blender, solve most clean shots and are where everyone should start. Resolve's Fusion and Cinema 4D add more control inside programs many artists already use. Mocha Pro handles the flat-surface and difficult-texture shots that point trackers struggle with. SynthEyes and PFTrack are dedicated matchmove programs for people who track a lot of shots. Nuke and 3DEqualizer sit at the top of the feature-film pipeline, where the matchmove department and the compositing department both have to be pixel-perfect on a cinema screen.

If you want to try this yourself, start small. Film ten seconds of a textured floor with plenty of detail while you walk slowly sideways, so there is lots of parallax. Use a fixed lens, avoid zooming, and keep the shutter fast enough to limit motion blur. Track it in After Effects if you already have it, or in Blender if you want it free. Set the ground and origin, drop a piece of 3D text on the floor, add a shadow catcher and some grain, and watch it stick. Once that works, try harder shots: a handheld walk, a shot with people crossing the frame, a wide lens with visible distortion. Every hard shot teaches you why the professional tools have all those extra buttons.

Sources and further reading

• Adobe, "Tracking 3D camera movement" (After Effects User Guide): helpx.adobe.com
• Blender Manual, "Motion Tracking" (Movie Clip Editor): docs.blender.org
• Foundry, Nuke Reference Guide, "CameraTracker": learn.foundry.com
• Blackmagic Design, DaVinci Resolve and Fusion product pages and reference manual: blackmagicdesign.com
• Maxon, Cinema 4D documentation, "Motion Tracker": help.maxon.net
• Makers' pages: Mocha Pro · SynthEyes · 3DEqualizer4 · PFTrack
• Tim Dobbert, *Matchmoving: The Invisible Art of Camera Tracking* (2nd ed., Sybex, 2012)
• Richard Hartley and Andrew Zisserman, *Multiple View Geometry in Computer Vision* (2nd ed., Cambridge University Press, 2004)
• Bill Triggs et al., "Bundle Adjustment: A Modern Synthesis", *Vision Algorithms: Theory and Practice* (Springer, 2000)

*All figures are illustrations drawn in your browser from one simple 3D scene, not real footage or screenshots of the software. The software windows are simplified drawings, not the real interfaces. Product details are the makers' own descriptions as of October 2026 and may change.*