1 The idea
The solve error number can’t tell you everything about a camera. The honest check is simple CG on the ground, such as a cone, with the shot playing: if the CG slides against the plate, the camera is wrong, whatever the number says.
Click to drop up to 4 cones on the ground, near and far, then switch between cameras A, B, C and D. Watch each cone’s slide now and worst, next to the camera’s average and worst-frame error. Notice which cameras fool a single far or centre cone.
2 Hands-on
Task: place one cone nearer than 12 m and one farther than 25 m, then find the camera where no cone slides more than 1.5 px.
Read every cone’s worst value, not just slide now, and play the whole shot. A camera that holds one cone can still slide another.
3 Deep dive
The error number isn’t the final word
Solve error tells you, on average, how many pixels the tracks sit from where the solved camera puts them. It doesn’t tell you whether the picture will look stuck: a solve with a low average can still slide or swim.
The real check is simple CG on the ground, such as a cone, with the shot playing. Eyes catch slide in motion, so a single still frame is not enough.
When the number and the picture disagree, trust the picture.
Cover it all: near, far, centre, edges
One cone in the middle of frame is easy to fool: several kinds of wrong camera still hold a single spot. Where the slide shows up tells you what is wrong.
| What you see | Likely cause |
|---|---|
| Far holds, near slides | A rotation-only (nodal) fit on a shot where the camera really moves |
| Centre holds, edges slide | Wrong focal, or missing lens distortion |
| Holds most of the shot, jumps on a few frames | Bad frames in the solve |
| Holds everywhere, all shot long | The camera is good |
Don’t edit the focal of a solved camera
Focal, lens distortion and the camera’s rotations are solved together as one answer. Change only the focal and the rest no longer fits: the picture slides more the farther you get from the centre. If you need another lens, re-solve with it.
A rotation-only fit has its own condition: it only holds when the move is really nodal. If the camera travels, the near ground slides.
A lens from the camera report is an input to the solve, not a fix to type in afterwards.
Go through the whole shot
A good average can hide a few bad frames. Always read the worst-frame error next to the average, play the whole shot, and step frame by frame where you have doubts.
To measure it, use Reconcile3D to turn a 3D ground point and the camera into a 2D track, then stabilise on it: the plate at that point should stand still.
- 1Cones near and far
- 2Play the whole shot
- 3Read worst, not just average
- 4Step frame by frame
- 5Stabilise with Reconcile3D
Check the layout too
A camera that doesn’t slide can still have a wrong layout. A big checkerboard Card as the ground should meet the plate’s horizon; a Cube with its pivot at the bottom should sit flat; an object of known size should read at the right scale.
4 In real work
In Nuke
- Take the solved Camera and a large Card with a CheckerBoard texture as the ground.
- Place small geo, such as a Cube with its pivot at the bottom, near, far and close to the frame edges.
- ScanlineRender through the Camera, then Merge (over) onto the plate (redistort the CG first if there is lens distortion, as in TR5).
- Play the whole shot, then step frame by frame where the worst-frame error is high.
- Reconcile3D: Camera + plate + an Axis at a ground point → create keyframes for a 2D track → stabilise and check the point stands still.
Common mistakes → the fix
- ✗ Checking with one cone in the middle of frame✓ Place cones near, far and close to the edges.
- ✗ Looking at a still frame instead of playing the shot✓ Play the whole shot, then step through the doubtful frames.
- ✗ Typing the “real” focal into a solved camera✓ Re-solve with that lens as an input.
- ✗ Accepting a solve because the average error is small✓ Read the worst-frame error and watch ground CG in motion.
Before you send the shot
- Ground CG holds both near and far.
- The edges hold as well as the centre.
- The whole shot has been played, and the high-error frames stepped through.
- Stabilised on a ground point with Reconcile3D, the plate stands still.
- The checkerboard ground meets the horizon and a known-size object reads at the right scale.
5 Check yourself
A far cone holds, but a cone close to the camera slides. What is the most likely cause?
🔒 Answer the question above and this card opens
1. A camera’s average error is low, but its worst-frame error is much higher. What do you do?
2. The solved focal doesn’t match the lens in the camera report. What do you do?
Where the numbers come from: A 48-frame handheld shot (640×360) rendered in Blender (Cycles) through our desert set, with a green screen on a pole and a drone flying across. Frame 1 is the still plate’s camera. Every true position comes from the scene’s own 3D (truth.json, dist.f32). Cameras A–D are built in your browser from the scene’s truth (A: the teaching solve from TR6 with good tracks; B: A with the focal changed to 30 mm and nothing re-solved; C: a rotation-only fit from frame 1’s position; D: A with three frames nudged). They are teaching cameras, each built to show one fault. Slide is measured against the true 3D ground point.