Agentic change detection: results so far

Snapshot of Nov 9, 2026. Blue bars with bold labels are ours, grey bars are baselines, hatched bars are ablations or upper bounds. Tap a chart for the full-size image. Charts marked partial are still running and will be updated.

Files: results.csv (every number in the charts) and the source tables: 3DGS-CD per scene, SceneDiff per pair, object results, scene update, edit ladder, robot pick-and-place, runtime on SceneDiff, PASLCD and 3DGS-CD.

Headline numbers (Nov 10)

Paper figures: object poses

Print size (3.5 in, 7 pt). SVG versions with editable text: pose_ladder.svg, pose_ablation.svg, pose_per_object.svg.

PSNR and depth error per edit step
Every scene with moved objects or parts (grey) and their mean (green): appearance against the held-out photo and depth against the after-change scan, per edit step.
Per-object depth error before vs after posing
Each moved object and part: depth error left at its old pose vs at our fitted pose. Below the diagonal = posing helped.
Detection ablations: from a bare agent to ours
Change detection: (a) from a bare agent to our method on 5 scenes, (b) one part of the protocol at a time on all 20, (c) judge model and a 4-minute scan. SVG
Pose ablations
Ablations of the pose step, scored where the moved objects and parts stand.

Ablations and baseline checks (tables)

Change detection

Mean F1 of the binary change masks at the evaluation views, each benchmark scored with its own script; IoU printed beside each bar. Every baseline is shown at its best labeled variant.

Change detection F1 on PASLCD, SceneDiff and 3DGS-CD
All four change-detection charts. CL-Splats on PASLCD uses a threshold picked on the test labels (upper bound). On SceneDiff, O-SCD's own pose step fails on 13 of 20 pairs, so it is also shown with our poses. 3DGS-CD's released labels mark only where objects moved to; on its held-out views 3DGS-CD still leads.

Object poses

The scene before the change is edited one kind of decision at a time. Poses are fitted on the even photos after the change and scored on the odd ones.

PSNR where all moved objects and parts stand, per edit step
All moved objects and parts together (rigid + articulated) at their fitted pose, 9 scenes.
PSNR where rigid movers stand, per edit step
Inside each rigid mover's outline at its fitted pose (no labels needed), 9 scenes.
PSNR where articulated parts are, per edit step
Doors and drawers at their fitted hinge angle or slide, 5 scenes.
PSNR inside the labeled change masks, per edit step
Inside the labeled change masks, all 20 instances (labels used only for this score).
Pose agreement between even-photo and all-photo fits
Fits from the even photos against fits from all photos, 22 moved objects. Both wrong ones are the same bowl in Cantina.
Median rigid fit IoU by dataset
Posed object outline against its place in the photos after the change, median per dataset.
Articulated joint fit IoU versus leaving the part in place
Articulated parts on PASLCD: fitted joint against leaving the part where it was, 15 parts.

Scene update

O-SCD's own update pipeline with its change masks and with ours, scored by its own script on held-out photos, full image.

PASLCD update before refit
PASLCD, 20 instances, before O-SCD's global refit.
PASLCD update after refit
PASLCD, after the refit.
CL-Splats update before refit
CL-Splats before the refit (partial).
CL-Splats update after refit
CL-Splats after the refit (partial).
Our object edit on the PASLCD odd split
Our object edit vs the unedited scene and O-SCD, PASLCD odd split (partial). Our refit adds no Gaussians; O-SCD's does and starts from its own 3DGS scene.
Our object edit on CL-Splats
Our object edit on CL-Splats (partial).

Simulated robot

Simulated pick-and-place successes per robot
The updated scenes as MuJoCo scenes; the robot carries each loose object from its place before the change to the place our method found. Success: held, final pose within 2 cm (ADD-S), robot never more than 3 mm into furniture.

Runtime

Runtime per SceneDiff pair
SceneDiff, median minutes per pair on one workstation GPU, queue waits excluded. Our method pays an upfront cost per scene; the case for it is quality. PASLCD and 3DGS-CD runtimes are in the tables above.