Skip to content

Tested Rigs & Datasets

LUC3D was developed and tested on recordings from the rigs below: 1 to 18 cameras, 30 to 150 fps, and 1 to 5 mice. Compare yours with them to see how close it is to something LUC3D already handles. Every rig recorded greyscale video of mice.

For parts, see What to Buy. For how many cameras you need and which calibration board to get, see Choosing Cameras & Calibration Boards.

At a glance

Rig Cameras Frame size, rate Video Animals Calibration board Mainly used to test
Home cage, 2022 8 1280×1024, 30 fps H.265 1–4 mice, mixed coats small hand-held ChArUco Track All identities against proofread tracks
18-camera arena, 2024 18 1680×1200, 120 fps H.265 2–3 mice printed 8×11 ChArUco, 24 mm Many cameras: speed, memory, which cameras matter
Open field, 2026 8 1280×1024, 60 fps H.265 2–5 mice 8×11 ChArUco, 24 mm; calib.io 14×9, 20 mm ID-switch checks, long recordings
High-speed arena, 2026 5 1280×1024, 150 fps H.264 with B-frames 2–3 mice not recorded 150 fps playback, lazy loading
Choice box, 2026 8 (7 calibrated) 1520×1040, 40 fps H.265 usually 2 mice in view not recorded campy recordings from another rig
Single top camera 1 1280×1024, 40 fps H.264 with B-frames 2 mice, light coats none Single-camera tracking
campy live-tracking rigs, 2026 3–4 1280×1024, 30 or 60 fps H.265 or H.264 3 mice calib.io 14×9, 20 mm campy recordings and live-tracking exports

The angles below are how far each camera looks down from horizontal, taken from its calibration (90° is straight down). Fields of view are horizontal.

8-camera home cage (2022)

  • Cameras: 8 greyscale cameras at 1280×1024, 30 fps. Two look almost straight down (80–86°), four are oblique (50–75°), and two are low side views (about 20°). Fields of view 67–81°.
  • Arena: a clear acrylic cage with bedding.
  • Video: re-encoded for proofreading to H.265 (CRF 12, denoised), about 10 minutes (18,000 frames) per session, a keyframe every 250 frames.
  • Animals: 1 to 4 mice. The ground-truth set is 22 proofread sessions: 6 with 3–4 mice of different coat colours (white, brown, black) and 16 with mostly two mice.
  • Calibration: a small hand-held ChArUco board; its size and dictionary were not recorded.
  • Useful for: scoring Track All against proofread identities in every camera, and tuning and checking the image and size ID-switch checks, the match gate and the tail-node weights. Its top and middle views also tested single-camera tracking. Read from a Google Drive folder, it showed that a cloud-synced file can read back zeros until it is downloaded.

18-camera arena (2024)

  • Cameras: 18 cameras at 1680×1200, 120 fps: 5 below the arena looking up through its clear floor, 9 around the sides and 4 above. Fields of view 57–83°.
  • Arena: a clear rectangular arena in an aluminium-extrusion frame.
  • Video: H.265 at constant QP 18 (about 52 Mbit/s per camera) or QP 24 (about 7.6 Mbit/s), a keyframe every 250 frames, 5–10 minutes per session.
  • Animals: 2–3 white mice, with a 34-node skeleton that includes the limbs and paws.
  • Calibration: a printed 8×11 ChArUco board, 24 mm squares, DICT_4X4_1000.
  • Useful for: LUC3D's speed and memory with many views: decoding only the views on screen, the frame cache budget, the timeline, and Triangulate All on a large eager project. It is also the rig behind the analysis of which cameras matter most (Choosing Cameras & Calibration Boards).

8-camera open field (2026)

  • Cameras: 8 greyscale cameras at 1280×1024, 60 fps. The layout changed between sessions: 1–2 cameras look (almost) straight down, and the others range from level to about 70°. Seven have a 75–80° field of view; one has a longer lens (43°).
  • Arena: a square open field with grid-patterned walls, inside a sound-attenuating enclosure.
  • Video: H.265 from an NVIDIA GPU encoder at constant QP 20–24 (about 3–4 Mbit/s per camera), no B-frames. Calibration videos at 5–10 fps.
  • Animals and sessions:
    • 5 white mice with 0–4 tail marks, 30 minutes (108,000 frames per camera). The tail marks give the true identities.
    • 2 white mice fighting, 10 minutes.
    • 3 white CD-1 mice, two of them in black mesh jackets, 15 minutes.
  • Calibration: an 8×11 ChArUco (24 mm squares, DICT_4X4_1000, the 18-camera rig's pattern) in March–April 2026, then the pattern of calib.io's coarse 300×200 mm board (14×9, 20 mm squares, DICT_5X5_1000) from May 2026.
  • Useful for: the ID-switch checks (identities checked against the tail marks), tracking with the tail nodes weighted 0, Track All's speed on a 108,000-frame lazy project, and flipped 2D poses that the match gate leaves ungrouped. campy's live 3D tracking was also validated on every 3-camera subset of two of these sessions (campy: Calibration).

5-camera high-speed arena (2026)

From another lab.

  • Cameras: 5 greyscale cameras at 1280×1024, 150 fps. One looks straight down (87–88°) and four oblique ones (53–69°) all look in from the same side. Fields of view 40–52°.
  • Arena: a square open field with light walls and a dark floor.
  • Video: H.264 with B-frames, about 2.6 Mbit/s per camera, 20 minutes (180,000 frames per camera).
  • Animals: 2–3 white mice.
  • Useful for: playback at 150 fps, lazy loading of long recordings, decoding video with B-frames, and the frame cache's effect on stepping backwards.

8-camera choice box (2026)

From another lab, recorded with campy.

  • Cameras: 8 FLIR Blackfly S BFS-U3-28S5M, cropped to 1520×1040, 40 fps, triggered on Line3 by an Arduino, 2 ms exposure. One camera looks straight down, the others are oblique (15–63°). Only 7 are in the calibration.
  • Arena: a clear two-chamber choice box with bedding.
  • Video: H.265 from an NVIDIA GPU encoder, about 3–4 Mbit/s per camera, 4–10 minutes per trial.
  • Animals: usually 2 mice in view.
  • Useful for: campy recording folders from another rig, an uncommon frame size, and a camera that has no calibration.

Single top camera (2023–2026)

  • Cameras: one camera above the cage, 1280×1024, 40 fps (the side videos were not used).
  • Video: H.264 with B-frames, 5–7 minutes per condition.
  • Animals: a resident mouse and a partner (BALB/c or CD-1); all light coats, so coat brightness cannot tell them apart. 207 proofread conditions, about 22 hours.
  • Calibration: none; there is no 3D with one camera.
  • Useful for: single-camera Track All, removing duplicate detections, and the single-camera review list.

campy live-tracking rigs (2026)

Rigs that record with campy and track in 3D while they record.

  • Cameras: 3–4 FLIR Blackfly S BFS-U3-13Y3M (1280×1024, mono), hardware-triggered on Line3 by an Arduino Nano ESP32, one exposure for every camera.
    • On a Windows PC: 30 fps, H.265 from Intel Quick Sync (quality 26, full range), 200 µs exposure, 7 dB gain, a home cage with 3 light-coated mice.
    • On a Linux mini-PC (GMKtec EVO-X2): 60 fps, H.264 from the AMD GPU (VA-API).
  • Calibration (Windows rig): calib.io's coarse 300×200 mm board (14×9, 20 mm squares, DICT_5X5_1000).
  • Useful for: reading campy's recording folders (session.json, the calibration check), animal identities from campy-live's .slp exports, hour-long projects (a 1.5 GB one-hour export, built by repeating a 30-second one), and a calibration whose first camera sits exactly at the world origin, as campy's calibrations put it.

Other calibration boards tested

A calibration test in July 2026 used calib.io's coarse 600×400 mm board (14×9, 40 mm squares, DICT_5X5_1000), the same pattern at twice the size of the 300×200 mm board above.