Telekinesis Agentic Skill Library

Appearance

CalibrationBenchmark ​

SUMMARY

CalibrationBenchmark cross-validates one target/IntrinsicOptions configuration: cross-validated reprojection error, plus an optional hand-eye round-trip error. Call run() once per configuration you want to compare, and collect the BenchmarkResults yourself.

The Skill ​

python
from telekinesis.axon import CalibrationBenchmark, IntrinsicOptions

The Code ​

python
from telekinesis.axon import CalibrationBenchmark, IntrinsicOptions
from telekinesis.axon.targets import CharucoTarget

target = CharucoTarget(squares_x=6, squares_y=9, square_length=0.012, marker_length=0.009)

bench = CalibrationBenchmark()
result = bench.run(
    images=my_images,
    target=target,
    options=IntrinsicOptions(),
    robot_poses=my_robot_poses,             # optional, enables the round-trip metric
    eye_in_hand_options=IntrinsicOptions(), # optional
    n_splits=5,
)
CalibrationBenchmark.save_json([result], "calibration_benchmark.json")
print(CalibrationBenchmark.format_results([result]))

Initialization ​

python
bench = CalibrationBenchmark()

CalibrationBenchmark takes no constructor parameters, every configuration is passed to run() instead.

Skills ​

SkillDescription
RunFull benchmark: cross-validated reprojection error, plus an optional hand-eye round-trip error.
Cross ValidateMean reprojection error over held-out folds of a set of images.
Hand-Eye RoundtripRecover base_T_board from all frames and report the round-trip reprojection error.

Best Practices ​

  • Compare held_out_reprojection_error to reprojection_error, not just look at either alone. They should be close. A held-out error that's substantially higher than the full-fit error means the model is overfitting the specific views it was calibrated on, usually a sign of too few or too repetitive views (see IntrinsicCalibrator Best Practices) rather than a fundamentally bad camera.
  • There's no universal "good" reprojection error, it depends on resolution and the precision your downstream task needs. Treat a benchmark run as a comparison tool: calibrate the same data with a couple of IntrinsicOptions variants (different corner detectors, different calibrate_camera_flags) and let held_out_reprojection_error pick between them, rather than chasing an absolute number.