The ICP benchmark currently registers the source against a synthetic near-identity transform (translation (0.05, -0.02, 0.01), 0.02 rad about z). That measures per-iteration speed only — it does not test registration robustness or accuracy.
The task
- Add a realistic target (e.g. a genuinely offset/rotated frame, or the next frame in a sequence).
- Report convergence and final alignment error, not just wall-clock time, alongside Open3D.
This makes the benchmark a fair accuracy comparison, not just a speed one. See the "Method notes" caveat in docs/benchmarks.md.
The ICP benchmark currently registers the source against a synthetic near-identity transform (translation (0.05, -0.02, 0.01), 0.02 rad about z). That measures per-iteration speed only — it does not test registration robustness or accuracy.
The task
This makes the benchmark a fair accuracy comparison, not just a speed one. See the "Method notes" caveat in docs/benchmarks.md.