Skip to main content
Back to timeline
arXivSource publication:

Sparse revisits days apart can still be replayed exactly: ReScene stores each visit as an immutable update and cuts history storage 2.5–4.8×

Synopsis

When a scene is revisited hours or days later, objects may move, disappear or deform, yet each revisit yields only a few images. ReScene stores each visit as an immutable compact update, so any earlier visit is recovered exactly by applying the updates in order without keeping past images; it attributes pixel disagreement between the prior reconstruction and the new images to individual Gaussians through their compositing weights, then suppresses outdated Gaussians, refines persistent ones and initialises newly visible surfaces from learned multi-view depth. Across Multi-State, WAT and ChronoScene it matches or exceeds 3DGS checkpointing quality with 2.5–4.8× less uncompressed history storage.

Source-provided article image: ReScene: Incremental Scene Chronology from Sparse Revisits
arXiv

Interpretation

Scene chronology no longer needs a complete model per visit: each revisit is stored as a compact update, and applying the updates in order recovers any earlier visit exactly without retaining past images. Fine-tuning and checkpointing at every visit duplicates the unchanged scene, so storage grows redundantly with each visit, while updating a single shared model alters earlier states. Across Multi-State, WAT and ChronoScene, reconstruction quality matches or exceeds sequential 3DGS checkpointing with 2.5–4.8× less uncompressed history storage, whereas sequentially updated ChronoGS loses 4.16–8.67 dB on earlier visits at final recall.

Updates land only on the Gaussians that actually changed, leaving unchanged surfaces largely untouched, so later revisits do not corrupt earlier states. Under sparse views, disagreement between the prior reconstruction and new images can come from scene change or from reconstruction error, and partial changes such as a bite taken from a tomato occupy only a small region, so inaccurate localisation either blocks needed updates or alters persistent surfaces. On the real Multi-State scenes, replacing the photometric change masks with CL-Splats DINOv2 masks or LTGS SAM masks lowers PSNR by 0.55 and 2.01 dB respectively, and applying CL-Splats projected-centre voting to its own masks instead of compositing-weight attribution lowers PSNR by a further 2.70 dB.

Newly exposed surfaces are initialised from learned depth that is consistent across views, rather than from triangulation that needs greater view overlap. Surfaces such as the interior revealed by opening a grill lid have no Gaussians in the prior model, and sparse revisits may not supply the view overlap triangulation requires. On the synthetic Multi-State scenes, at arrival PSNR within 0.64 dB of sequential ChronoGS, foreground depth error is 44.59 mm versus 144.19 mm for sequential ChronoGS, about one third.

Perspective

The demonstrated setting allows up to 24 training images per revisit, with the scene static within each visit and changing between visits; evaluation covers four real and four synthetic Multi-State scenes, ten WAT scenes and twelve ChronoScene scenes, with ground-truth geometry at every state on the synthetic scenes and offline-estimated, cross-visit-aligned camera poses on the real scenes.

When the evidence guiding an update is unreliable, ghosting artefacts may appear: occlusion and appearance variation can corrupt change support and misguide suppression, and depth errors can misplace births so they duplicate existing surfaces. The ablations are single-seed comparisons describing observed differences rather than statistically established effects, and refinement is not uniformly beneficial: on the Plant scene, removing it raises PSNR from 30.07 to 31.04 dB.