Skip to content

Choosing Cameras & Calibration Boards

Two choices to make before you build a rig: which calibration board to buy, and how many cameras to buy and where to put them. This page covers both. To record and compute the calibration itself, see Calibrating Cameras.

Calibration boards

calibrat3 needs a ChArUco board: a chessboard with an ArUco marker in every white square. Every measurement LUC3D makes in 3D inherits the board's errors, so buy a board made for calibration rather than printing one yourself.

What makes a good board

Property Why it matters Rule of thumb
Rigid and flat The solver assumes all corners lie on one plane. A bowed board moves corners off it. Aluminium composite, ceramic or glass, never paper on cardboard. calib.io advises laser- or inkjet-printed targets only for testing
Big enough A board that fills more of each image pins down focal length and distortion better Held face-on, it should cover about half of each camera's image (calib.io). It still has to move and tilt freely inside your arena
Squares big enough Small squares are detected less reliably, especially when the board is tilted At least 20 px across in the farthest camera (calib.io), more if you tilt the board a lot
Coarse pattern, not Fine Fine boards have squares half the size, so about four times as many corners Coarse. calib.io suggests Fine for cameras above about 3 MP, but in the lab's tests Fine was never clearly more accurate and failed more often (below)
Matte, evenly lit Glare and shadows hide markers Diffuse light; no point sources reflecting off the board

To check the square size before you buy:

square width in pixels ≈ square size (mm) × focal length (px) ÷ distance from the camera (mm)

Read the focal length in pixels from an existing calibration.toml (matrix[0][0]), or work it out as lens focal length ÷ pixel size: an 8 mm lens on a sensor with 3.45 µm pixels is 8 ÷ 0.00345 ≈ 2,320 px.

The 18-camera rig

Its cameras sit 34–60 cm from the arena centre, which gives 2.2–3.9 px per mm there. 40 mm squares (calib.io Coarse 600 × 400, the board recommended for it below) come out at 88–156 px, and 20 mm squares (Coarse 300 × 200) at 43–78 px, both comfortably above 20 px. 10 mm squares (Fine 300 × 200) come out at 22–39 px: just enough face-on, too small once tilted.

calib.io boards

calib.io sells ChArUco boards on 6 mm aluminium/LDPE composite, UV-printed with a matte finish, made to order. All use the DICT_5X5 marker dictionary and an odd number of rows. The lab uses the Coarse 300 × 200, 400 × 300 and 600 × 400 boards.

Board (mm) Coarse: rows × columns, square / marker Fine: rows × columns, square / marker Price (USD)
200 × 150 9 × 12, 15 / 11 mm 17 × 24, 7.5 / 5.5 mm $132
300 × 200 9 × 14, 20 / 15 mm 17 × 28, 10 / 7 mm $137
400 × 300 9 × 12, 30 / 22 mm 17 × 24, 15 / 11 mm $153
600 × 400 9 × 14, 40 / 30 mm 17 × 28, 20 / 15 mm $231
800 × 600 9 × 12, 60 / 45 mm 17 × 24, 30 / 22 mm $311

For very close cameras (a few centimetres), calib.io also sells Fine ceramic boards of 75 × 50, 100 × 75 and 150 × 100 mm (2.5–5 mm squares, $501–$1,041). Prices are from calib.io's US store on 29 October 2026.

Which board for which rig

Your rig Board
Small: one cage, cameras within about 30 cm (such as the 4-camera ChronicRecord rig) 300 × 200 Coarse. The 200 × 150 Coarse only if the 300 × 200 cannot move freely in the cage
Medium: an arena of about 0.5–1 m (such as the 8-camera arena rig) 300 × 200 or 400 × 300 Coarse
Large: many cameras around a big space (the 18-camera rig, marmoset cages) 600 × 400 Coarse. The 800 × 600 was no better

This comes from a lab study that re-analysed 15 calibration recordings from nine rigs and tried every calib.io board in simulated copies of three of them:

  • Size matters little. On real recordings the 300 × 200 Coarse gave 0.12% error on lengths the calibration had not seen (0.3 mm over 24 cm), on both ChronicRecord and the 8-camera arena. In the simulated rigs, boards of different sizes were within a few hundredths of a millimetre of each other, so pick the size that moves freely through your space.
  • No Fine boards. Fine boards were at best 0.02 mm more accurate, and only once. They could not be read from as far away: the 7.5 mm squares of the Fine 200 × 150 were unreadable from ChronicRecord's farther cameras. 4 of 12 calibrations with Fine boards failed in sleap-anipose's solver, against none of 20 with Coarse or printed boards.
  • The recording matters more than the board. Record 45–60 s that brings the board through the whole space in front of every camera, with synchronized cameras and a flat board. On the real rigs, the calibrations that failed came from short recordings (about 30 s) in which some cameras rarely saw the board.
  • The board comparison is simulated so far. No real recording yet compares boards on the same rig in one sitting.

Entering the board in calibrat3

Enter a calib.io board in calibrat3's Board panel like this:

Field Value
Squares X the columns from the table (e.g. 14)
Squares Y the rows (e.g. 9)
Square (mm) / Marker (mm) from the table, then check them with calipers on your board
Dictionary 5x5 (1000)
Pattern auto (calibrat3 tries both OpenCV corner orderings)

Other vendors

Vendor What they sell Notes
calib.io (Denmark) Standard ChArUco boards, above; custom patterns on request Made to order
JD Photodata (UK) Inkjet-printed ChArUco with the rows and columns you choose, on 3 mm Dibond (aluminium composite), 100 × 150 to 400 × 600 mm From about US$101. High-resolution print: square pitch within ±0.1 mm; a measurement certificate is extra. Inkjet, which calib.io advises only for testing
Foamcore Print (USA) Custom ChArUco on 3 or 6 mm aluminium composite (also foam or acrylic), from 4 × 4 in up to 48 × 96 in Manufacturing tolerance ±0.004 in (0.1 mm). Priced per configuration. An option for boards bigger than 800 × 600 mm
Print your own calib.io's pattern generator (internal or educational use) or OpenCV Print without scaling, mount it flat on glass or aluminium composite, and measure the printed squares. Good for trying a layout; buy a manufactured board for real calibrations
Edmund Optics Chrome-on-glass checkerboards Very accurate, but small and without ArUco markers, so calibrat3 cannot use them

How many cameras, and where

Every camera adds cost (the camera, lens, cable, a USB port, disk space) and makes every processing step slower. To find out which cameras earn their place, we took one recording from an 18-camera rig and switched cameras off in software.

In short

  • Spread cameras over heights instead of adding more at one height. Nine cameras split between bottom and top put 77% of body points within 5 mm of the full rig; nine side cameras put 60%.
  • Each height sees different body parts. Bottom cameras see paws and legs best; top and side cameras see the head best.
  • The side cameras overlap the most, so they are the first to cut. Top cameras were the most valuable per camera.
  • Returns diminish fast. The best 8 cameras get 85% of body points within 5 mm, the best 12 get 93%.
  • More cameras do not rule out identity swaps. Some sets of 9–14 cameras swapped the two mice once where the full rig did not, so plan to proofread.

The test rig

Height Cameras Where Sees best
Bottom 5 25–35 cm below the clear floor, looking up Paws, legs
Low side 5 Around the arena, from 15 cm below to just above floor level, looking up at it at 10–35° —
Upper side 4 Around the arena, 15–20 cm above the floor, looking across Head
Top 4 40–65 cm above the floor, looking down Head

The recording: two mice in a glass arena (they used about 34 × 21 cm of floor), 2 minutes at 120 fps (14,401 frames), 1680 × 1200 cameras, SLEAP predictions on a 34-node skeleton (not proofread). The cameras sit 34–60 cm from the arena centre.

How it was measured
  1. Track All with all 18 cameras grouped each mouse's detections across views.
  2. Every body point on every 8th frame (117,112 points) was triangulated from all 18 cameras, the reference, and again from each camera set. Both used DLT, dropping the worst view while any view was more than 15 px off (like Triangulate's reprojection-error threshold).
  3. Score: the share of reference points that a camera set puts within 5 mm of the reference. A point that fewer than two of the set's cameras see counts as a miss.
  4. Identities: Track All was re-run with only that set's cameras (the others excluded, as in the Tracking Wizard's camera list), and each detection's identity compared with the full rig's. A flip is a stretch of at least 30 frames where the two mice's identities are the other way round.
  5. For scale: two separate halves of the rig (9 cameras each, mixed heights) disagree with each other by a median 3.7 mm. The full rig is the best estimate available, not ground truth.

Results by height

Cameras kept Number Body points within 5 mm Median distance Paws within 5 mm Head within 5 mm Identity flips
Bottom only 5 53% 4.2 mm 70% 42% 0
Side only (low + upper) 9 60% 3.0 mm 58% 81% 2
Top only 4 40% 5.6 mm 23% 69% 2
Bottom + top 9 77% 2.3 mm 76% 83% 1
Low side + top 9 78% 1.8 mm 67% 91% 2
Bottom + low side 10 77% 2.1 mm 80% 83% 2
All but top 14 84% 1.3 mm 85% 92% 0
All but bottom 13 87% 1.0 mm 80% 97% 1
All but low side 13 91% 0.9 mm 89% 96% 0
All but upper side 14 93% 0.8 mm 91% 96% 1
All 18 18 reference reference

What this says:

  • No height works alone. The bottom cameras miss much of the head (they could triangulate only 74% of head points at all); the top cameras miss the paws.
  • Leaving out a whole height costs, per camera saved: top 4 points, bottom 2.6, low side 1.8, upper side 1.8. Keep the top and bottom cameras; thin out the side rings first.
  • A bottom view needs a clear floor. Without one, put more cameras low on the sides.

The best set for each budget

Share of body points within 5 mm of the full 18-camera rig, by number of cameras The best set of each size rises quickly to 70% at 4 cameras and 85% at 8, then flattens. Random sets of the same size are much worse. Single heights (bottom only, side only, top only) score 35 to 60%. 0% 25% 50% 75% 100% 2 4 6 8 10 12 14 16 18 Number of cameras Body points within 5 mm 15 random camera sets of each size: from 4%–53% at 3 cameras to 97%–100% at 17 Best 2 cameras: 45% of body points within 5 mm Best 3 cameras: 59% of body points within 5 mm Best 4 cameras: 70% of body points within 5 mm Best 5 cameras: 75% of body points within 5 mm Best 6 cameras: 79% of body points within 5 mm Best 7 cameras: 82% of body points within 5 mm Best 8 cameras: 85% of body points within 5 mm Best 9 cameras: 87% of body points within 5 mm Best 10 cameras: 89% of body points within 5 mm Best 11 cameras: 91% of body points within 5 mm Best 12 cameras: 93% of body points within 5 mm Best 13 cameras: 94% of body points within 5 mm Best 14 cameras: 96% of body points within 5 mm Best 15 cameras: 97% of body points within 5 mm Best 16 cameras: 98% of body points within 5 mm Best 17 cameras: 100% of body points within 5 mm Best 18 cameras: 100% of body points within 5 mm Bottom only (5 cameras): 53% of body points within 5 mm Bottom only Side only (9 cameras): 60% of body points within 5 mm Side only Top only (4 cameras): 40% of body points within 5 mm Top only Bottom + top (9 cameras): 77% of body points within 5 mm Bottom + top Low side + top (9 cameras): 78% of body points within 5 mm Bottom + low side (10 cameras): 77% of body points within 5 mm All but top (14 cameras): 84% of body points within 5 mm All but bottom (13 cameras): 87% of body points within 5 mm All but low side (13 cameras): 91% of body points within 5 mm All but upper side (14 cameras): 93% of body points within 5 mm Best set Random sets Whole heights
Share of body points within 5 mm of the full 18-camera rig. Line: the best set of each size. Shaded: the range of 15 random sets of each size. Diamonds: whole heights switched on or off, as in the table above. Hover a point for its value.
Cameras Bottom Low side Upper side Top Body points within 5 mm Median distance Identity flips
4 2 – 1 1 70% 2.4 mm 2
6 2 1 1 2 79% 1.8 mm 0
8 3 2 1 2 85% 1.5 mm 0
10 3 2 2 3 89% 1.1 mm 0
12 3 3 2 4 93% 0.8 mm 0
18 5 5 4 4 reference reference

The sets were found by removing, one at a time, the camera whose loss hurt the score least. For 8 cameras the table lists a set with two low-side cameras instead of a second upper-side one: it scores the same (84.6% against 84.7%) and kept the identities. Where the cameras go matters as much as how many: 15 random sets of 8 cameras scored 68–82% (median 79%, the same as the best 6), and random sets of 4 scored 22–66%.

How to use this when buying:

  • 4 cameras: 70% of body points within 5 mm if they are spread out (two below, one side, one top), but with two mice the identities flipped twice.
  • 6–8 cameras: the best value on this rig: 79–85% of body points within 5 mm and no identity flips in this recording. Put two or three below (if the floor is clear), two on top, and the rest around the sides at different heights.
  • 12 cameras: 93%. The last 6 cameras of the 18 add only 7 points.

Limits of this test

One rig and one 2-minute recording of two white mice, with SLEAP predictions that were not proofread. The reference is the full rig, so a set is judged by how close it gets to what all 18 cameras say. The identity flips come from a couple of hard moments when the mice were close together and few cameras saw one of them. At the first, following each mouse in 3D shows the full rig was right and the side-only set swapped the mice; without proofread labels, the others cannot be scored. Copy the pattern (spread over heights, a few cameras below and above), not the camera numbers, which depend on this rig.