Proof first · then try it

See how LiDAR map ghosts are reduced.

A moving scene can look clean scan-by-scan and still accumulate ghost geometry. This gallery separates the audited detector-free proof from visual demos, so you can check the evidence before choosing a path.

⚡ Interactive · No install

Try the real NumPy library in your browser

Choose Box, Range, or Temporal, run an Argoverse 2 or nuScenes preset, or drop your own LiDAR scan (.pcd / .bin / .xyz / .npy). No GPU, no upload, no signup. This is a visual preview; the audited metrics are below.

Strict GT Proof

Detector-free AV2 map audit

1,235,563 points are evaluated against 84,471 moving-track GT points. Raw, cleaned, and TP/FP views use the same pose-aligned frame; configuration and confusion counts are saved as JSON.

66.3% moving GT removedprecision 65.1% / F1 65.7%

97.4% static GT keptboxes are used only to create the evaluation labels

Visual demo · Box-driven

AV2 accumulated sequence

A 20-frame public-data preview using per-frame annotation boxes and pose alignment. It shows how transient objects contaminate an accumulation; its removed-point count is not a detector-free moving-GT score.

20-frame accumulated map See how moving-object contamination builds up in the map.

Per-frame boxes + pose alignment Reproducible from public AV2 data and annotations.

Visual demo · Local

Local sequence comparison

A checked-in multi-frame visual comparison using temporal consistency. Useful for understanding the interaction, not a replacement for the audited benchmark.

Fast check

Single-scan preview

The shortest visual check: remove points from one point cloud and inspect the result in 3D.