3D GAUSSIAN SPLATTING · SCENE CHRONOLOGY

ReScene

Incremental Scene Chronology
from Sparse Revisits

Reconstruct a changing scene.
Return to any observed state.

Reconstruct change. Keep the history.

The can changes shape across four visits. ReScene retains a reconstruction of each observed state.

01 · LOCALISE CHANGE

Evidence for each Gaussian.

Compare the preceding reconstruction with the new observation. ReScene attributes change and stability evidence to the Gaussians that contribute to each pixel.

Tomato, previous reconstruction
Previous reconstruction
Tomato, new observation
New observation
Tomato, projected change weight
Projected change weight
Low change weightHigh change weight

This example shows the first update of Tomato from a selected training viewpoint. Purple indicates a low change weight and yellow a high change weight. Reconstruction errors and appearance differences can also produce a response, so the weights are evidence of image disagreement rather than labels of physical change.

02 · RETRIEVE A VISIT

Return to an earlier state.

All updates are complete. Select a visit to see its stored reconstruction beside the held-out observation.

Held-out observation, Can visit 1
Held-out observation
ReScene final recall, Can visit 1
ReScene · final recall

Can · Visit 1 of 4 · recovered after the final update

THE MULTI-STATE DATASET

Non-rigid changes.
Distinct observed states.

Synthetic Multi-State

Controlled changes with known geometry

Four synthetic scenes cover folding, soft deformation, articulated motion and moving leaves. Each video completes an orbit before switching to the next observed state.

Cloth

Folding and reopening

Teddy

Bending, twisting and compression

Plant

Branch and leaf motion

Robot

Articulation and payload transport

Real Multi-State

Objects captured in changing configurations

Four real objects show deformation, articulation, material removal and changes in object configuration. Each video shows the captured object at distinct visits.

Tissue

Tissue extraction and packet changes

Tomato

Progressive material removal

Can

Opening and crushing

Gimbal

Changes to the gimbal and its cover

MAIN RESULTS

Reconstruction quality and history storage

Results from the main paper across real and synthetic Multi-State, WAT and ChronoScene.

Multi-State · Real · PSNR (dB), SSIM and mean history storage per scene (MiB).
MethodPSNR ↑SSIM ↑Storage ↓
CL-Splats21.030.82691.2
Seq. ChronoGS (arrival)24.920.863108.6
Seq. ChronoGS (recall)18.900.774
Seq. LTGS21.870.844304.7
Seq. 3DGS24.020.861410.5
Ours25.400.873163.6

SELECTED COMPARISONS

Detail that survives later updates.

After the last update, each method reconstructs the earlier visits. The camera moves within each state, and cuts separate the observed visits.

Earlier states remain available.

The tissue box changes between visits. ReScene recovers its earlier reconstructions after the final update.

Reconstruct each change in shape.

The cloth folds and reopens across ten observed states. ReScene reconstructs each state without estimating the motion between visits.

THE METHOD

Update selectively. Retain each state.

ReScene reconstructs a changing scene from sparse images collected at successive visits. It attributes image disagreement to individual Gaussians, updates the current reconstruction and stores each visit as an immutable update.

ReScene workflow: compare the preceding reconstruction with new images, attribute change and stability evidence to Gaussians, and use evidence and estimated depth to refine, suppress and add Gaussians. Optimise the update against the observations.
The ReScene workflow.
01

Gather evidence

Compare the previous reconstruction with the new images. Accumulate change and stability evidence using each Gaussian’s contribution to the rendered pixels. Estimated depth supports surface placement and refinement.

02

Update the reconstruction

Reduce the opacity of outdated Gaussians, refine selected positions, scales and colours, and add Gaussians for missing surfaces. Optimise these updates together using the current images.

03

Retain the history

Store the updated attributes alongside the initial model. Recover an earlier visit by applying its sequence of updates. Later updates leave that stored reconstruction unchanged, including any errors it contains.