Trajectory Accuracy (ATE / RPE)

Accuracy is measured with the two standard VIO/SLAM trajectory metrics, computed clean-room from their definitions (no ov_eval/evo source contact), and driven by an EuRoC replay harness (#46).

The metrics (branes::sdk::eval)

Absolute Trajectory Error (ATE)

ATE rigidly aligns the estimated trajectory to ground truth — VIO has a gauge freedom, so the global pose is arbitrary — and reports the translational RMSE of the aligned positions. The alignment is Horn’s closed-form absolute-orientation solution (Horn, 1987): the optimal rotation is the eigenvector of the largest eigenvalue of a 4×4 profile matrix.

Implementation note worth keeping: the profile matrix’s top two eigenvalues are routinely near-degenerate, which makes power iteration converge far too slowly (and silently return a wrong rotation). The implementation uses a cyclic Jacobi eigensolver on the 4×4 instead — robust regardless of the eigenvalue gap.

Relative Pose Error (RPE)

RPE compares relative motions over a fixed step Δ: for each i, it compares Pᵢ⁻¹·Pᵢ₊Δ between estimate and ground truth and reports the translational RMSE of the difference. RPE is invariant to a global rigid transform, so it measures local drift rather than global alignment.

A associate() helper matches estimated poses to the nearest-in-time ground-truth pose within a tolerance.

The EuRoC replay harness (branes::sdk::euroc)

Parsers read the ASL CSV layout — imu0/data.csv (gyro+accel), cam0/data.csv (image timestamps + filenames), state_groundtruth_estimate0/data.csv (T_world_body) — and a replay driver feeds a VioEstimator the time-ordered IMU+image stream (loading PNGs on demand), returning the estimated trajectory. Parsers reject non-finite values and zero-norm quaternions; a single unreadable frame is skipped rather than aborting the run.

How it runs in CI vs. locally

The metrics and CSV parsers are unit-tested in CI on synthetic fixtures (ATE = 0 for identical / rigidly-transformed trajectories, ≈ 0.1 m for a known offset; RPE = 0 under a global transform; association windowing; ASL parsing). The full V1_01_easy replay + ATE-threshold assertion is dataset-gated — set CORTEX_EUROC_V101=/path/to/V1_01_easy/mav0 to run it; otherwise it skips so ctest -R vio_euroc stays green without the ~1.5 GB sequence.

The accuracy gate

The gate below is pinned to the test threshold (regenerated from the source on every build):

Sequence	Gate (ATE)	Reference (published, keyframe)
EuRoC V1_01_easy	< 0.5 m	OpenVINS ~0.05–0.06 m · VINS-Fusion ~0.08 m

The gate is deliberately generous relative to tuned SOTA: this MVP MSCKF has no online intrinsics/extrinsics refinement and no loop closure, so the bar is set to catch a broken pipeline without over-fitting to a tuned number, and is documented in the test file to be tightened as the estimator improves.