Result formats

What a run writes, and which file to use for what.

Dual YAML — camera calibration

Camera calibration writes two files on purpose:

File	Contains	Use for
*_calib.yaml	Intrinsics, distortion, extrinsics, time offsets, readout	Deploy. The calibration itself
*_full.yaml	The above plus per-frame board poses	Replay. Re-run analysis without redetecting

Deploy the small one; keep the full one to reproduce the run.

Both carry a timings: block — hierarchical per-stage timings with metadata, so you can ask whether a change was faster overall or only in one stage.

output_format: json writes JSON instead.

Cam-IMU JSON

camcalib cam-imu writes a JSON result containing, per sensor:

Field	Meaning
T_CtoI	Camera → IMU transform
td	Camera–IMU time offset
readout	Rolling-shutter readout, when estimated
Dw, Da, Tg	IMU intrinsics, in the configured convention
b_g, b_a	Biases
t_ga	Gyro↔accel offset, when estimated
gyro_only	Per-IMU flag — the result self-documents which IMUs skipped the accel
source topic	Per sensor; this is what lets apply-calib derive the topic→sensor map without a hand-written file

Intrinsics are convention-dependent

Dw / Da are only comparable between calibrations that used the same imu_intrinsics_model. Read the convention before comparing numbers — see IMU model.

Batch layout

With input.bags, each sequence gets its own discoverable result:

<output_dir>/
  bag00/cam_imu.json
  bag01/cam_imu.json
  bag02/cam_imu.json

That layout is what the repeatability study reads.

HTML report

One self-contained file per run — no external assets, so it can be emailed. Contents: navigable TOC, summary stat cards, intrinsics and distortion tables, extrinsics with an inline 3D plot, temporal calibration (time offset and readout together), reprojection analysis (per-frame robust median and P95, per-tag, error-vector histograms, dual-panel scatter, corner heatmaps), calibration-data quality (motion-excitation metrics with grades, rotation-axis spread, angular-velocity timeline, feature-coverage heatmap), detection statistics, pipeline timing, and collapsible appendices. Print-friendly CSS.

Rebuild reports from saved YAML plus the detection cache, without recalibrating:

python -m camcalib.analysis.regenerate_html

Kalibr interop

camcalib.io.kalibr is the one surface for Kalibr files:

Function	Direction	Handles
read_camchain	in	Intrinsics + distortion across pinhole / ds / eucm / fov
read_camchain_imucam	in	Also each camera’s T_cam_imu and stereo T_cn_cnm1
write_camchain_imucam	out	Exports a cam-IMU result as a round-trippable camchain-imucam

This is how a Kalibr calibration becomes a CamCalib starting point, and how a CamCalib result goes back into a Kalibr-expecting pipeline.

Ground-truth comparison

Per-dataset ground-truth YAMLs live under examples/regression/gt/. Where one exists, the dashboard annotates each parameter with its deviation — for example Δ +0.27% GT=190.97 — colour-coded by accuracy band.

Corrected bags

camcalib apply-calib writes a corrected ROS1 bag plus a provenance sidecar recording which calibration was applied and what it changed. The bag is ordinary — anything that read the original reads the corrected one.