Calibench: Leaderboard Results

Machine-readable backend of the Calibench leaderboard — one JSON record per (tool, dataset, split) evaluation, plus the calibration outputs they came from. Browse it on the Space; this repo is the data of record.

Layout

tracks.json                      # track registry: dataset rosters per (track, split),
                                 #   metric config {primary, columns}, scoring tau_max
datasets.json                    # dataset catalog: per-rig hardware, shutter, lens,
                                 #   IMUs, board, source/license/cite
schemas/result.schema.json       # record schema (v1.1)
<Tool>/<batch>/<id>.json         # result records   (batch = evaluation run date)
<Tool>/<batch>/<id>.output.json  # the calibration outputs (intrinsics, extrinsics,
                                 #   time offsets) each record's calibration_ref points at

Batches group the records produced by one evaluation run (YYYY-MM-DD), so a leaderboard state traces to an exact run set and the same tool can be compared across time. The leaderboard ranks each tool's latest batch per dataset by default; its Compare batches toggle splits methods into tool @ batch.

Verification: verification.status is self_reported until the Calibench CI re-runs the evaluation (in progress) — self-reported reprojection statistics are not directly comparable across tools (each reports its own statistic); the leaderboard flags such mixed rankings.

Citing a leaderboard state

Cite the commit hash shown in the leaderboard status line (every view names the exact revision of this repo it was rendered from), plus the Calibench citation in calibench-vi/datasets.

Adding results

Open a PR to yangyulin/calibench-submissions with your calibration outputs; CI validates against the schema and publishes records here. Records (the numbers/metadata) are CC-BY-4.0.