acceptodds
Under review as a conference paper at ICLR 2027

Learning from Automatic 3D Gaussian Splatting Edits for Online Scene Change Detection

Abstract

Recent advances in online scene change detection relocalise incoming query images against a reference 3D Gaussian Splatting (3DGS) model. Changes are then detected by fusing photometric and foundation-model cues from each query and its pose-matched reference rendering. Per-image normalisation aligns the numerical ranges of different cues for fusion, but makes response strength relative to each image, limiting comparability across views. We address this limitation through scene-specific learning from automatic 3DGS edits. Using only the reference 3DGS, this fully automatic process selects object instances, modifies their appearance or geometry, and renders labelled changed and unchanged examples for training. As query images arrive, a causal generalised Bayesian formulation combines the learned prior with current image evidence, current-view geometric support, and historical support from the previous Gaussian field state. The field is updated through direct evidence accumulation, without iterative gradient optimisation or replay of past frames. Evaluation on PASLCD shows that our pixel pipeline achieves mIoU and F1, outperforming O-SCD Online by mIoU points. It reduces full-pipeline computation time by an estimated and change inference time by a measured using the same image cues. Object-guided mask refinement raises performance to mIoU and F1 while remaining online, exceeding the reported scores of GS-Diff, which requires post-change reconstruction, by mIoU points and F1 points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Change detection, Gaussian Splatting, Computer Vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MkVoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles online scene change detection (OSCD) against a reference 3D Gaussian Splatting (3DGS) model, aiming to overcome the cross-view incomparability inherent to per-image normalization. To calibrate cue fusion without manual supervision, the authors propose an automated data synthesis pipeline that partitions reference Gaussians using SAM2 prompts, applies synthetic edits (recoloring, translation, removal), and fits five scene-specific calibration and prior modules offline. At runtime, streaming query images are processed via a closed-form local Bayes factor formulation that accumulates change evidence into the 3D Gaussian field causally and without gradient backpropagation, followed by an optional connected-component SAM2 refinement step. Empirical validation on the PASLCD benchmark indicates that the basic pixel pipeline attains 60.2 mIoU and 73.6 F1 (reported as a 38.9x inference acceleration over O-SCD Online), while the full refinement pipeline reaches 66.5 mIoU and 78.1 F1, outperforming offline reconstruction baselines like GS-Diff.

Critical strengths

S1: The formulation of online scene change detection through a localized Bayes Factor update provides an elegant mathematical framing. Deriving a closed-form analytic solution for evidence accumulation avoids online gradient backpropagation, parameter fine-tuning, or recurrent neural state tracking during streaming inference. Decoupling the per-frame feature discrepancy update from heavy online iterative optimization is theoretically sound and establishes a clean probabilistic foundation for accumulating change beliefs directly within the 3D Gaussian Splatting representation.

S2: The proposal to synthetically generate labeled change samples directly on the reference 3D Gaussian model addresses a genuine data scarcity challenge in 3D change detection. By leveraging 2D foundational segmentation priors (SAM2) to partition the reference radiance field and applying parametric transformations (color interpolation, opacity attenuation, and spatial rigid displacements), the framework constructs a self-contained surrogate dataset without requiring manual paired ground-truth annotations across physical scene captures.

S3: The empirical results on the PASLCD benchmark demonstrate strong overall performance, and the authors provide extensive ablation studies breaking down the modular pipeline. The experimental validation systematically probes the individual contributions of the suppression classifier, isotonic regression calibration, and contour refinement, illustrating that each layer incrementally improves the resulting Intersection-over-Union metrics.

Critical weaknesses

W1: The narrative trajectory in the introduction exhibits a significant conceptual leap that compromises its motivation. The core problem highlighted in existing literature is that per-image normalization disrupts cross-view metric comparability. However, the proposed solution bypasses standard statistical calibration or global feature normalization entirely, instead prescribing a cumbersome, scene-specific synthetic data generation and offline multi-stage model fitting pipeline. The necessity of jumping from "inconsistent per-image scale" to "synthesize edits on reference Gaussians and train five separate discriminative/regression modules per scene" is neither intuitively justified nor contrasted against lightweight baseline alternatives (such as global photometric calibration, canonical cross-view whitening, or multi-view geometric variance scaling).

W2: There is a critical disconnect between the speed claims and the accuracy claims throughout the abstract and introduction. The authors proclaim a 38.9x speedup over O-SCD alongside a 66.5 mIoU score that surpasses offline reconstruction methods such as GS-Diff. In reality, these two results originate from two distinct pipelines evaluated in isolation: the 38.9x throughput figure is achieved exclusively by the bare pixel pipeline (which scores a markedly lower 60.2 mIoU), whereas the competitive 66.5 mIoU necessitates the full pipeline with per-frame, multi-connected-component SAM2 decoder passes for contour refinement (Equations 8–9). Mask decoder invocations in streaming loops incur non-trivial computational overhead. Presenting the highest accuracy alongside the highest speedup creates an asymmetric attribution that obscures the true Pareto frontier of the complete system.

W3: The temporal Bayesian evidence accumulation formulated in Equation (6) violates the standard assumption of temporal conditional independence. In robotic streaming and handheld camera trajectories, sequential frames possess extreme viewpoint correlation, redundant visual observations, and shared systematic artifacts (such as local 3DGS reconstruction blur, high-frequency specular floaters, or micro-calibration drift). Directly summing the log Bayes factors log⁡Btg\log B_{tg} across consecutive frames without an observation discount factor, temporal decay parameter, or spatial covariance modeling causes rapid, uncalibrated confidence accumulation. In scenarios where a camera hovers or moves slowly across a viewpoint with slight rendering artifacts, this accumulation will compound minor static discrepancies into aggressive false-positive detections.

W4: The synthetic edit generation paradigm suffers from an inherent domain gap regarding novel object additions. The transformation generator operates strictly on the existing primitives of the reference scene via color shifting, alpha scaling (removal), and rigid spatial displacement. It cannot synthesize novel, unobserved geometric topologies, out-of-distribution surface textures, or non-rigid deformations that characterize external objects brought into a real-world environment. Because the response suppressor TT and the log-likelihood ratio mapper ℓimg\ell_{img} are fitted exclusively on variations of the reference scene's native Gaussians, their generalization to genuinely novel visual classes and external clutter remains theoretically unsupported and vulnerable to out-of-distribution classification failures.

W5: The paper advertises a real-time, online causal paradigm while withholding the substantial computational and wall-clock overhead incurred by the scene-specific offline initialization. Setting up the system on a new reference scene requires prompting SAM2 across uniform grid points, clustering 3D Gaussians, rendering multi-view augmented trajectories, and sequentially fitting five distinct components: the response suppressor, the isotonic regression model, the beta prior statistics, the pixel prior, and the contour filter. Treating this extensive preparation as negligible offline pre-processing conflicts with real-world deployments where systems must adapt on-the-fly to newly mapped environments.

W6: The evaluation scope is heavily concentrated on the PASLCD benchmark, where the target change distributions (primarily object displacement and removal) align closely with the authors' hand-crafted transformation heuristics. The paper lacks stress-testing under complex real-world perturbations, such as dynamic lighting transitions (e.g., diurnal illumination shifts or abrupt shadow casting), non-rigid structural deformations (e.g., fabrics, curtains, plant foliage), and scenarios involving severe viewpoint occlusion boundaries where 3DGS depth discontinuities produce boundary artifacts.

Questions for authors

Q1: What is the exact end-to-end wall-clock time breakdown required to prepare a single reference scene before online inference can commence on standard hardware (e.g., an NVIDIA RTX 3090/4090 or A100)? Specifically, how many minutes are consumed by (a) SAM2 prompting and 3D Gaussian clustering, (b) rendering the transformed trajectories, and (c) fitting the five respective regression and classification modules?

Q2: What is the true end-to-end runtime performance (measured in latency per frame in milliseconds, throughput in FPS, and peak GPU memory utilization) of the Full Pipeline when online SAM2 contour refinement (Equations 8–9) is actively running on every detected connected component? How does this latency scale as the number of candidate change instances in a frame increases?

Q3: How does the cumulative log-evidence update in Equation (6) behave when the camera remains static or moves along a very slow, looping trajectory over regions with minor view-dependent reconstruction flaws? Have the authors experimentally validated a temporal discounting factor or an exponential moving average update to prevent runaway false-positive evidence accumulation?

Q4: Could the authors provide a fine-grained quantitative ablation on the PASLCD benchmark decomposed by the ground-truth change typology—specifically differentiating performance on "Newly Added External Objects" versus "Removed Existing Objects" and "Displaced Objects"? Does the model exhibit a noticeable drop in recall or precision on entirely new object categories that do not share Gaussian priors with the reference model?

Q5: In Equations (4) and (5), the scale parameter of the Half-Normal distribution is fixed to δ∼HalfNormal(1)\delta \sim \text{HalfNormal}(1), and spatial neighborhood covariance is approximated with a scalar parameter cc. What is the sensitivity of the final mIoU to variations in these hyperparameters, and how are these constants determined for diverse scenes with radically different spatial extents and spatial Gaussian density distributions?

6 comments
AuthorsOpenReview ↗

Reference preparation, motivation, and calibration (W1/W5/Q1)

Thank you for recognizing our closed-form Bayesian updates, automatic reference-edit supervision, and strong PASLCD results. We failed to update the OpenReview abstract to match the submitted PDF, causing confusion about our motivation (W1) and accuracy–speed trade-offs (W2). It also omitted the percent signs: "38.9" referred to a 38.9% reduction in change-inference time (approximately 1.64× speedup), not 38.9×. The abstract now matches the revised manuscript. We apologize for these errors and the resulting confusion.

Why use offline reference editing for online detection?

The interval between reference capture and later queries allows offline preparation, motivating methods that use 3DGS to render reference views for change detection. We further use the reference to estimate statistics for closed-form Bayesian updates and generate edits to supervise response modelling and output refinement. Each query is compared with its reference rendering without online optimization or past-image replay. Although this preparation takes place before online detection, we agree that it remains a computational burden, which is now also acknowledged in the limitations section. More details are provided in §1, "Introduction".

Why use LLR mapping given the suggested calibration alternatives?

We appreciate your suggestions for alternative calibration strategies. Reference editing provides supervision for suppression, LLR mapping, pixel-prior fitting, and output refinement. Given this supervision, a monotone LLR mapping is a simple way to convert response strength into change evidence. Per-image normalization now only motivates this choice; the broader motivation is to learn from the reference scene for accurate and efficient online detection. More details are provided in §§4.2–4.3.

How long does each preparation stage take on standard hardware? Across the four datasets, average times are 212.0–431.5 seconds on RTX 5090 (Ryzen 7 9800X3D), 240.1–481.0 seconds on RTX 4090 (Threadripper 9960X), and 292.0–598.0 seconds on RTX 3090 (Threadripper PRO 9955WX), including model loading and file I/O. The following table reports PASLCD means over 20 instances on all three hardware systems.

Mean preparation time (s/instance), given an existing reference 3DGS:

StageRTX 3090RTX 4090RTX 5090
Preprocessing11.19.28.0
SAM2 masks and 3D association108.175.760.0
Data generation53.056.252.8
Response suppression27.121.318.1
LLR mapping25.020.916.9
Unchanged-reference statistics5.06.810.0
Pixel prior10.08.59.0
Contour filter45.335.631.2
Initialization and other overhead7.67.99.9
Total292.0242.0215.8

More details are provided in Appendix G.11, "Additional Hardware Measurements".

Can preparation be reduced for faster deployment?

The revised manuscript compares three configurations, Base, Standard, and Full, to quantify the accuracy benefits and computational costs of the added components (as we responded on Component contributions and practical trade-offs.). Base requires no reference edits; its offline preparation is estimated at 27.9–51.0 seconds per scene instance on RTX 5090, including preprocessing and remaining overhead. Even Base exceeds online O-SCD in mIoU and F1 on all four datasets. Reference-edit learning and refinement together improve Full over Base by 7.3–18.7 mIoU points across four datasets. More details are provided in §5, "Runtime".

AuthorsOpenReview ↗

Accuracy, latency, and memory (W2/Q2)

What are the latency and FPS of the complete pipeline?

With the original hyperparameter settings unchanged, the updated PASLCD runtime measurements on RTX 5090 give 27.7 FPS for Standard and 17.5 FPS for Full, compared with 25.1 and 16.6 FPS, respectively, in the original submission.

Timings cover each query from input loading to the final mask.

Online latency (ms/frame) and FPS; mIoU (%).

PASLCD

Config.PoseRenderR0Infer.RefineTotalFPSmIoU
O-SCD online25.20.65.837.7—69.314.449.24
Base26.10.65.84.4—37.027.057.40
Standard25.10.65.64.8—36.127.760.13
Full25.20.65.65.320.357.217.567.08

CCS3D

Config.PoseRenderR0Infer.RefineTotalFPSmIoU
O-SCD online31.02.35.7177.0—216.04.649.94
Base30.92.15.912.5—51.619.455.05
Standard30.32.05.613.1—51.019.666.67
Full31.02.05.613.015.867.514.873.77

CL-Splats

Config.PoseRenderR0Infer.RefineTotalFPSmIoU
O-SCD online33.81.85.7174.8—216.24.661.57
Base34.01.65.712.3—53.718.663.85
Standard33.41.75.712.5—53.318.867.32
Full34.31.75.712.717.471.813.971.12

3DGS-CD

Config.PoseRenderR0Infer.RefineTotalFPSmIoU
O-SCD online53.81.46.1100.5—161.96.267.14
Base49.51.36.09.0—65.815.277.75
Standard48.71.35.99.5—65.415.377.41
Full49.31.35.99.718.885.011.888.09

Base and Standard run approximately 1.9–4.2× as fast as online O-SCD across the four datasets. Full achieves higher accuracy while retaining a 1.2–3.2× speedup, including output refinement. More details are provided in Table 3B.

How does latency grow with candidate count? To examine the cost of processing more candidate regions, we group all 500 PASLCD queries by their number of connected components before small-component filtering on RTX 5090.

Refinement latency (ms/frame).

Candidates per frameFramesMedian (ms)P95 (ms)
0–167515.030.8
17–3214616.218.4
33–6412818.535.9
65–1289822.730.4
129–2564129.342.7
>2561261.6108.6

Median refinement latency increases with candidate count, with substantially higher costs above 256 candidates. Image features are shared across regions, while additional candidates require more decoder batches and mask processing. More details are provided in Appendix G.10, "Candidate count and memory".

What is the peak GPU memory of the complete pipeline? Separate raw-image-to-final-mask profiles on seven PASLCD instances record maximum PyTorch allocated/reserved peaks of 3.03/5.13 GiB, including localization, rendering, feature extraction, and refinement. More details are provided in Appendix G.11, "Memory usage of Full".

AuthorsOpenReview ↗

Bayesian assumptions and parameter scales (W3/Q3/Q5)

Correlated observations and evidence accumulation (W3/Q3)

Does Eq. (6) violate temporal conditional independence?

No. Eq. (6) defines an accumulated evidence field rather than a joint likelihood, and the loss-based generalized Bayesian update in Appendix D uses this field without assuming conditionally independent frames. We agree, however, that correlated frames can still affect accumulation in practice, which we examine below.

How does our existing design account for potential error accumulation?

Our design already includes the spatial covariance modeling you mention: the reference-estimated coefficient cc adjusts the conditional mean and variance of each local Bayes factor (Eq. (12)), and setting c=0c=0 lowers Full mIoU by 4.37 points. Historical increments can also be negative, allowing unchanged observations to reduce accumulated evidence. We evaluate discounting and temporal decay below.

Following your suggestions, we evaluate discounting and EMA on the original PASLCD sequences and under repeated observations, and additionally examine sparse observations.

Do discounting and EMA improve the original sequences? Yes, but the benefit depends strongly on the parameter setting. Discounting at H=32H=32 improves Full mIoU by approximately 0.07 points; the other 11 settings reduce it.

Both rules use the retention factor λ=2−1/H\lambda=2^{-1/H}, where HH counts effective observations of a Gaussian. Discounting scales previous evidence by λ\lambda; EMA additionally scales incoming evidence by 1−λ1-\lambda. We evaluate both rules at H=1,2,4,8,16,32H=1,2,4,8,16,32 across all 20 PASLCD instances, with corresponding reference-fitted priors and contour filters. All 12 settings reduce FPR, but 11 also reduce recall.

Full-output metrics (%). Recall is the true-positive rate (TPR). Metrics are averaged over frames within each instance, then equally across the 20 instances.

UpdateHalf-life HRetention λmIoU ↑FPR ↓Recall ↑
Original update (Table 1)—1.000067.080.73477.97
Discounting10.500065.980.62175.44
Discounting20.707166.450.63976.28
Discounting40.840966.860.67277.08
Discounting80.917066.950.71177.57
Discounting160.957667.010.72377.79
Discounting320.978667.150.72677.98
EMA10.500065.680.61575.00
EMA20.707165.540.62074.99
EMA40.840965.260.63474.78
EMA80.917065.030.62374.35
EMA160.957664.670.62273.92
EMA320.978664.320.61773.53

More details are provided in Appendix G.6, "Original sequences".

Half-Normal scale and spatial co-response (Q5)

How is cc determined across scenes, and how sensitive are the results? cc is estimated from each scene’s unchanged reference observations using Eq. (12) of the original manuscript, so its value can vary across scenes. On PASLCD, setting c=0c=0 removes the co-response correction and reduces Full mIoU from 67.08% to 62.71%, a drop of 4.37 points.

How is the Half-Normal scale determined, and how sensitive are the results? We first standardize responses relative to their unchanged-reference level and uncertainty. In this standardized space, τ=1\tau=1 provides a natural unit-scale default for the non-negative response increase. We use this setting across datasets and assess its sensitivity below. Testing τ=0.5,1,2\tau=0.5,1,2 yields 66.89%, 67.08%, and 66.86% Full mIoU, respectively, on all 20 PASLCD instances, indicating low sensitivity within this range on this dataset. More details are provided in Appendix G.7, "Bayes-factor sensitivity and controls".

These comparisons support our modelling choices, but do not establish robustness to substantially different scene extents or Gaussian densities. As acknowledged in the limitations of the original manuscript, the response uncertainty and spatial covariance models underlying our local Bayes factors are working approximations.

Reviewer_2xdAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the online scene change detection tasks by identifying changes in query images against a reference 3DGS representation of a scene. The query images are matched to the reference using XFeat keypoints and pose is estimated using PnP-RANSAC, in order to compare it with a render of the reference at that pose. The output of the proposed method is a pixel-level binary change map, produced via a combination of models fitted per scene for evidence accumulation using automatic edits. The offline stage encompasses the fitting of a logistic response-suppressor, log-likelihood-ratio, unchanged-response, pixel prior and the contour filter. The online stage includes pose estimation, utilizing pixel, SSIM and feature maps for differencing, evidence splatting and re-rendering, prediction as a log-odds sum of prior, current evidence and history using neighboring gaussians and finally mask refinement using SAM2. Regarding contributions, while editing images or the 3DGS representations to produce before and after pairs as a training set is not novel (CYWS, CYWS-3Dm Zanjani et al.), using edits to calibrate a pixel prior and to fit a change detector per scene without labels is a new approach according to my knowledge. Furthermore, the authors approach of modelling the neighbourhood residual response as change evidence is an interesting approach under the Bayes inference formulation.

Validation and experimentation includes the PASLCD and CCS3D datasets, reporting mIoU, F1 and FPS numbers across real and synthetic scenes as well as ablation studies on PASLCD.

Critical strengths

The paper’s main strength is the per-scene pixel prior π fitted using automatic edits of the reference 3DGS representation. This component produces before and after pairs requiring no additional data nor manual annotation, and according to the ablation studies removing the pixel prior reduces mIoU by 16.47 meaning it’s the most contributing factor to the proposed approach results.

Furthermore, this component could be generalized as automatic edits could be utilized to calibrate other render-and-compare SCD methods, meaning it could be incorporated as a module for existing and future work.

Critical weaknesses

Aside from the novelty of the pixel prior, the proposed method relies mostly on components from O-SCD and 3DGS-CD including XFeat and PnP-RANSAC pose estimation, L1, D-SSIM and SAM2 feature maps, SAM2 mask refinement etc.

The validation presented in paper is lacking, covering two datasets with 14 scenes in total (PASLCD includes 20 instances, two for each scene, while CCS3D includes four synthetic scenes).Furthermore, all ablations and runtime results are on PASLCD only. Both real and simulated datasets are available: CL-Splats, NERFCD, 3DGS-CD data and others.

The proposed method includes hand-set hyperparameters which are not ablated or tested for sensitivity in many components such as the SAM2’s object proposals and cross-view tracking, edit generation, edit filtering, Gaussian graph, prior fitting and others. The prevalence of these constants alongside the small size of the test set allude to iterating on PASLCD to report the highest results. There are missing ablations, including the neighbourhood residual response, number of reference images in the 3DGS representation, number of query images, quality of the 3DGS representation reconstruction, number of edits performed on the reference 3DGS and hyperparameter sensitivity tests.

The limitation chapter is brief and does not adequately discuss the proposed methods limits and failure cases, including but not limited to change detection between image pairs, change-type prediction (added, removed, moved, etc.) and object-level change detection. While some examples are present in the supplementary material, a broader inspection is required, particularly when the validation was performed on a small test set.

The paper is missing discussion about related work, including image pair (pairwise) change detection, multi-label semantic change detection, object-level change detection, as well as NeRF-based and 3D reconstruction based approaches, such as ChangeSim, C-3PO, CSCDNet etc. For example, relevant works that should be further discussed and compared against include SceneDiff, GOLDILOCS and 3DGS-CD (see list below).

Sakurada, K., Shibuya, M., Wang, W.: Weakly supervised silhouette-based se- mantic scene change detection. In: IEEE International conference on robotics and automation (ICRA). pp. 6861–6867. IEEE (2020)

Park, J.M., Jang, J.H., Yoo, S.M., Lee, S.K., Kim, U.H., Kim, J.H.: Changesim: Towards end-to-end online scene change detection in industrial indoor environments. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8578–8585. IEEE (2021)

Wang, G.H., Gao, B.B., Wang, C.: How to reduce change detection to semantic segmentation. Pattern Recognition 138, 109384 (2023)

Wu, Yuqun and Lin, Chih-hao and Che, Henry and Tiwari, Aditi and Zou, Chuhang and Wang, Shenlong and Hoiem, Derek, SceneDiff: A Benchmark and Method for Multiview Object Change Detection, ECCV 2026

Friedlander, Almog and Shamir, Ariel and Fried, Ohad, GOLDILOCS: General Object-Level Detection and Labeling of Changes in Scenes, ICLR 2026

Questions for authors

• How were the constants selected? Including edit strengths, IoU gates, filter thresholds, s_tg, τ, 8-NN and the 5:1 split. Were these chosen while iterating over PASLCD results? • How does the proposed method perform on additional datasets? Given frozen hyperparameters, report metrics on unseen real scenes such as 3DGS-CD. • How much of the gain can be attributed to Bayesian accumulation? Historical support adds 2.7 mIoU while π contributes 16.5 and the refinement stage adds about 7. Can you provide results for online O-SCD with (1) Per-scene threshold fitted by your self-edit procedure, and (2) the same object-guided refinement?

4 comments
AuthorsOpenReview ↗

Contributions and cross-dataset evaluation (W1/W2/Q2/Q3)

Thank you for recognizing the novelty of using reference edits to fit a pixel prior and a scene-specific change detector without manual labels, and for finding our Bayesian formulation of neighbourhood residual responses interesting. Your comments helped us clarify these contributions, examine component interactions, and broaden the evaluation.

Clarifying contributions and component interactions (W1/Q3)

How does our method differ from O-SCD? Although we reuse only O-SCD's localization and initial-response computation, we introduce recursive Bayesian updates that accumulate evidence without online optimization or past-image replay, reducing online computation. Our Bayesian Base configuration already exceeds online O-SCD in mIoU, F1 and FPS on all four datasets. Reference-edit supervision further supports response suppression, LLR mapping, prior fitting, and output refinement, enabling Full to achieve state-of-the-art online accuracy while remaining faster than online O-SCD.

How does our method differ from 3DGS-CD? The similarity is the use of a SAM decoder; how we use its outputs is fundamentally different. 3DGS-CD uses its outputs for cross-view object association/ mask fusion, followed by joint camera and object pose optimization. To avoid such online cost, our reference-edit-trained contour filter guides object completion for each query, without online optimization or past-image replay. We recognize that the above distinction might not have been sufficiently clear in the original manuscript and have clarified the refinement description. More details are provided in §4.2.

How much does Bayesian accumulation contribute beyond the prior? In the progressive ablation, aggregated support StS_t adds 1.2 mIoU points on PASLCD, and historical support HtH_t adds another 3.8 points, giving Bayesian fusion a total gain of 5.0 points over the initial response without reference-edit supervision. Base exceeds online O-SCD on all four datasets. The large prior gain reflects its interaction with the preceding components. More details are provided in Table 2.

O-SCD with our threshold fitting and refinement (Q3)

The gains vary across datasets, and O-SCD remains well below Full. In short, we tested reference-edit threshold fitting and our object-guided refinement separately and together; with both applied, the gap is 10.2–18.3 mIoU points across the four datasets. More details are provided in our response on threshold fitting and output refinement for O-SCD.

Performance transfer with fixed hyperparameters (W2/Q2)

Does the method generalize to additional datasets with fixed hyperparameters?

Yes. Full achieves state-of-the-art online detection accuracy on all four datasets with the original hyperparameter settings fixed. Please see our response on Expanded evaluation Component ablations and runtime breakdowns now cover all four datasets. More details are provided in Tables 1–3.

AuthorsOpenReview ↗

Parameter of Object proposals, tracking, and reference editing (W3/Q1)

Parameter selection and sensitivity (W3/Q1)

We acknowledge that the original manuscript did not sufficiently explain several parameter choices. These settings were not selected by repeatedly tuning against PASLCD test results. They follow standard defaults, statistical conventions, or empirical choices motivated by their intended roles. The sensitivity tests reported below cover all 20 PASLCD instances.

How were SAM2 proposals and cross-view tracking configured? We use SAM2’s default proposal-quality thresholds and a two-view (K=2K=2) tracking window. Each reference view is matched with the next two views (K=2K=2). These local matches can form longer tracks through intermediate views. The Benjamini–Hochberg procedure applies a strict FDR level of 0.01 to screen out chance overlaps between candidate masks.

We evaluate the tracking window on all 20 PASLCD instances, keeping FDR fixed.

Tracking windowFull mIoU (%)
K = 166.22
K = 2 (default)67.08
K = 467.33

Doubling the tracking window from two to four views yields only a 0.25-point mIoU gain while increasing cross-view association costs. Using one subsequent view reduces mIoU by 0.86 points.

How did we choose the reference-editing settings? We chose moderate recolouring and translation strengths to generate clear changes while retaining usable supervision. Recolouring uses an empirical mixing weight of 0.45mg0.45m_g, where mgm_g is the Gaussian's object membership, to distinguish edits from rendering discrepancies without concentrating supervision on only pronounced changes. Translation starts at 1.51.5 times the object extent to separate the original and moved regions while keeping both visible in the same view. A translation is accepted if, in at least one view, the moved object remains visible and overlaps its original mask by less than 1%. Removal represents a presence/absence change and therefore has no separate strength parameter. Edit-strength sensitivity: We jointly halve or double recolouring and translation strengths, keeping removal and filtering unchanged and reusing the default results.

Edit strengthRetained translation editsFull precisionFull recallFull mIoU
Half1184.2373.8764.06
Default48484.8578.2967.08
Double37187.7273.7665.79

With halved strength, most translations fail the separation checks, leaving almost no translation supervision alongside weaker recolouring and resulting in lower mIoU. With doubled edit strength, Full precision increases while recall decreases, indicating more conservative detection. A possible explanation is that fitting the prior on stronger synthetic changes may favour a decision point that misses weaker real changes, reducing mIoU. These results support the moderate default strength among the three tested settings.

AuthorsOpenReview ↗

Evidence updates, quality controls, and data requirements (W3/Q1)

How did we choose the evidence-update settings?

For evidence updates, τ=1\tau=1 gives a unit-scale Half-Normal prior for standardized response increases. We empirically use k=8k=8 for a small local neighbourhood and a 5:15{:}1 reference-view split to allocate most views to edit-based learning while reserving separate unchanged views for reference statistics.

SettingTested valuesMaximum Full mIoU deviation (points)
Half-Normal scale (τ)0.5, 1, 20.22
Neighbourhood size (k)4, 8, 160.11
Reference-view allocation2:1, 5:1, 11:10.12

Performance varies little over these tested ranges, indicating low sensitivity to the precise parameter values.

Does prior fitting require manual parameter tuning?

No additional scene-specific tuning is used. The pixel prior is estimated by maximizing reliability-weighted foreground IoU on reference edits, with reliability weights computed from reference-response statistics. More details are provided in §4.3, "Pixel prior fitting".

Do the quality-control thresholds require precise tuning?

No, within the tested ranges: changing these thresholds affects Full mIoU by at most 0.26 points. Object IoU gates screen poorly matched reference objects, while edit-filtering thresholds screen synthetic edits with excessive effects outside the labelled regions or insufficient valid background after removal. We set these as empirical quality-control criteria.

We shift both object IoU gates by ±0.10 and jointly halve or double the five edit-filtering limits:

Quality-control settingTested valuesMaximum Full mIoU deviation (points)
Mean/minimum object IoU(0.55, 0.40), (0.65, 0.50), (0.75, 0.60)0.16
Five edit-filtering limits0.5×, 1×, 2× defaults0.26

Strict and loose edit filtering yield 67.34% and 67.15% Full mIoU, respectively, compared with the default 67.08%. More details are provided in Appendix A.1 for parameter definitions and Appendix G.7 for sensitivity results.

Neighbourhood evidence and data requirements (W3/Q1)

To address the remaining ablation requests in W3, we then evaluate the neighbourhood residual response, observation gate, reference-image count, query-image count, reconstruction training budget, and edit budget on all 20 PASLCD instances. To assess the effect of reconstruction quality, we vary the number of 3DGS training iterations (7k, 15k, and 30k).

The following tests cover all 20 PASLCD instances.

Control or experimentComparisonFull mIoU (%)
Mechanism controls
Spatial co-response coefficient cEstimated → 067.08 → 62.71
Observation gatestg≥1s_{tg}\geq1 → stg>0s_{tg}>067.08 → 65.89
Budget and observation experiments
Edit budgetFull → half67.08 → 66.19
Reconstruction iterations7k / 15k / 30k65.98 / 66.76 / 67.16
Reconstruction training imagesAll / half / quarter67.16 / 65.15 / 59.35
Query countFirst 1 / 5 / 25 queries62.72 / 64.34 / 67.08

Does the neighbourhood residual correction help? Yes. Setting c=0c=0 removes the expected unchanged neighbourhood-response correction and reduces mIoU by 4.4 points on PASLCD.

Why set the observation gate to stg≥1s_{tg}\geq1? This requires at least one pixel-equivalent of accumulated rendering contribution, providing a minimum level of image support before updating a Gaussian's evidence. Relaxing the gate to stg>0s_{tg}>0 admits weakly supported observations and reduces Full mIoU from 67.08% to 65.89% on PASLCD.

How do data and reconstruction budgets affect accuracy? Halving the edit budget and shortening reconstruction training have modest effects, while reducing reconstruction views causes a larger drop, highlighting the importance of reference-view coverage. From 1 to 25 queries, Full mIoU increases from 62.72% to 67.08%, while the across-instance standard deviation decreases from 20.25 to 10.80 percentage points (O-SCD: 11.66 to 7.24). More details are provided in Appendix G.7, "Parameter Sensitivity and Reference Supervision".

AuthorsOpenReview ↗

Task scope, failure analysis, and related work (W4/W5)

Task scope and failure analysis (W4)

What about image pairs, change types, and object-level detection?

These are extensions beyond our current focus on accurate, low-latency binary detection with a reference 3DGS. We agree that the original limitations discussion was incomplete and now identify these extensions, which require further adaptation, as future work. More details are provided in §6, "Limitations and Conclusion".

What about the failure cases?

Localization failure. The original XFeat–MNN frontend in O-SCD fails on Sill in 3DGS-CD dataset at query 24. Replacing the matcher with LighterGlue, while retaining XFeat and the original pose-estimation procedure, enables both our method and O-SCD to process all 30 queries. This increases median localization time from 29.7 to 53.4 ms in separate-run diagnostics.

Historical accumulation can degrade detection. When a new object is inserted, history could reinforce incorrect associations when the same sparse viewpoints recur without intermediate observations that could correct them. In Sill, the inserted object's change evidence is associated with different background surfaces across these viewpoints, producing persistent false positives. This scene is the main contributor to the 3DGS-CD decline from 79.04% to 77.75% mIoU when adding historical accumulation. However, refinement corrects many of the resulting detection errors, improving Sill mIoU from 62.64% to 85.12%. More details are provided in §5, "Failure cases".

Relation to existing approaches (W5)

Thank you for recommending these works. We have now added CSCDNet, ChangeSim, C-3PO, SceneDiff, and GOLDILOCS, and expanded the discussion of 3DGS-CD and NeRF-based approaches in Related Work. We compare their input and reconstruction requirements, prediction targets, and computation during detection, clarifying how our emphasis on online binary-mask prediction differs from semantic and object-level change detection. These revisions distinguish the components we reuse from our evidence modelling, reference-edit learning, and recursive update framework. The revised manuscript also includes quantitative comparisons with SceneDiff and 3DGS-CD. More details are provided in §2, "Related Work", and Table 1.

Reviewer_22dVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a scene-specific approach to online scene change detection that assumes an existing reference 3DGS reconstruction. The method automatically edits the reference 3DGS to generate labeled training data. Such a data pipeline enables offline learning of scene-specific change priors and evidence models. During online inference, it combines current-view and historical evidence through closed-form Bayesian updates to avoid heavy iterative optimization and past-frame replay. The resulting pixel-level change masks are further refined using SAM2-based object-guided segmentation.

Experiments on PASLCD and CCS3D demonstrate improved detection accuracy of their method and efficient online inference compared to the O-SCD baseline. The paper also provides detailed ablations to distinguish the contributions of the core change detection pipeline from those of SAM2 mask refinement.

Critical strengths
  1. Efficient. The online frame-by-frame change detection is quite efficient without heavy iterative optimization or past-frame replay. The offline learning and calibration cost is also reasonable, taking only approximately 3.6 minutes per scene ( given an existing 3DGS reconstruction ).

  2. Strong performance. The qualitative results in the supplementary video are convincing, and the method demonstrates substantial quantitative improvements over the O-SCD baseline on both PASLCD and CCS3D. Importantly, the core change detection pipeline already achieves notable improvements without SAM2-based refinement.

Critical weaknesses
  1. The method relies on a pre-reconstructed reference 3DGS for each new scene. Obtaining a high-quality reference 3DGS can also be costly, which is not included in the reported offline preparation time.

  2. I would like to see more discussion about the practical applications of online scene change detection in the introduction. It's a useful task for scene representation and dynamic perception.

Questions for authors
  1. Would applying the same SAM2 refinement to O-SCD further improve its performance too? I am curious how much of the performance gap would remain under the same refinement setting. Also, how sensitive is the final performance to the choice of segmentation model used for refinement (e.g., SAM2 vs. SAM3)?

  2. The reported 3.6-minute offline preparation time assumes an existing reference 3DGS. How much time would the complete pipeline typically take, including reference image collection, 3DGS reconstruction, and offline learning/calibration? This would help assess the practicality of deployment in unseen environments.

2 comments
AuthorsOpenReview ↗

Applications, reconstruction, and preparation costs (W1/W2/Q2)

Thank you for recognizing the efficiency of our online updates and the accuracy gains achieved before output refinement. Your suggestions helped us clarify the deployment setting and evaluate the associated computational trade-offs.

Practical applications (W2)

Thank you for this suggestion, which helped us clarify the practical context and motivation. The revised Introduction discusses urban monitoring, construction inspection, and long-term robotic mapping, with supporting references. It then explains that reference observations precede subsequent queries, allowing reference processing and reconstruction beforehand. This provides the context for our further use of the reference scene to support efficient online detection. More details are provided in §1, "Introduction".

Reconstruction and preparation cost (W1/Q2)

This deployment setting separates preparation from online detection, but does not eliminate its cost. Following other reviewers' requests for broader evaluation, we expanded our experiments to four benchmarks covering 34 scene instances.

What is the total computational preparation cost? The preparation-cost table below reports mean wall-clock seconds per instance on an RTX 5090. Our preparation includes reference processing, data generation, and model fitting.

Mean time (s/instance):

DatasetCOLMAPFastGS reconstructionOur preparation
PASLCD35.579.9215.8
CCS3D34.582.7252.2
CL-Splats39.850.7212.0
3DGS-CD59.958.8431.5

Reconstruction and our preparation were measured separately. Summing these stages gives an approximate computational budget of 5.5 minutes for PASLCD, starting from available reference images. More details are provided in Appendix G.11, "Runtime and Offline Preparation".

How long does reference-image collection take?

PASLCD does not report reference-image acquisition times. Acquisition depends on scene size and capture requirements; for scenes of PASLCD's scale, we estimate that collecting high-quality reference images with a professional camera typically takes a few minutes in addition to computation.

How do reconstruction training and reference-view budgets affect detection?

We examine both training iterations and the number of reconstruction views on all 20 PASLCD instances, using FastGS with the supplied COLMAP poses and sparse initialization. The 7k, 15k, and 30k checkpoints come from the same training trajectory. At 30k iterations, we additionally train with every second or fourth reference image in filename order, retaining half or one quarter of the views. All reference images remain available for edit generation and fitting, isolating the reconstruction training-view budget. Each reconstruction receives matched reference editing and fitting before query evaluation.

Training time is the mean training-loop time per instance, excluding initialization; reduced-view timings are not reported.

FastGS training iterationsReconstruction training viewsTraining time (s/instance)Full mIoU (%)
7kAll14.065.98
15kAll33.066.76
30kAll55.067.16
30kHalf—65.15
30kQuarter—59.35

Reducing training from 30k to 7k iterations lowers mIoU by only 1.18 points, while halving or quartering the reconstruction views lowers it by 2.01 or 7.81 points, respectively. These results suggest that shorter reconstruction training can retain useful detection accuracy, while sufficient reference-view coverage is more important in this comparison. More details are provided in Appendix G.7, "Reconstruction quality and training-view count".

What are the preparation costs of the three configurations?

As defined in our response on component contributions and practical trade-offs, the three configurations are Base, Standard, and Full. Given an existing reference reconstruction, Base takes an estimated 27.9–51.0 s for preparation on RTX 5090, including preprocessing and remaining overhead. Standard takes an estimated 184.7–410.1 s for preparation while maintaining similar online latency, and Full takes 212.0–431.5 s, adding output refinement for higher accuracy. More details are provided in §5, "Runtime".

AuthorsOpenReview ↗

Refinement and segmentation-model choice (Q1)

Refinement and segmentation-model choice (Q1)

Does the same refinement improve O-SCD? It improves mIoU on two datasets but lowers it on the other two; O-SCD remains below our Full configuration on all four. More details are provided in our response on threshold fitting and output refinement for O-SCD.

How sensitive is accuracy to the refinement model? Our SAM2 implementation uses the SAM 2.1 Hiera-Tiny checkpoint. We also evaluated alternative segmentation models on the same 500 PASLCD queries while fixing the pixel pipeline. Latency below measures the refinement stage, not the complete detection pipeline.

Refinement modelFull mIoU (%)Refinement latency (ms/frame)
SAM267.0820.9
SAM 367.9687.5
EdgeTAM65.2418.4

SAM 3 improves mIoU by 0.88 points at 4.2 times the refinement latency; EdgeTAM is faster but less accurate. These results support the accuracy–latency trade-off of SAM2 used in our method. More details are provided in §5, "Runtime".

Across the four benchmarks, our method achieves state-of-the-art online scene change detection performance, supported by the expanded accuracy comparisons, component ablations, and runtime breakdowns.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to All Reviewers (1/2): Main Revisions and Evaluation

We thank the reviewers for their careful reading and constructive feedback. Their comments helped us recognize where the original manuscript lacked clarity and prompted more thorough experiments to better explain our motivation, design, and how the components contribute and interact. Below, we summarize the main revisions and findings shared across the reviews, followed by detailed responses to each reviewer in indivual comment.

Guide to highlighted revisions. To keep the revised manuscript readable, we highlight major revisions selectively in blue. In the main text, we mark section titles where substantial rewriting occurred, together with key passages addressing the reviewers' concerns. In the appendix, we mark only the titles of added or substantially revised sections.

Revised motivation (MkVo: W1/W5; 22dV: W2).
We revised the 'Introduction' to first introduce practical applications of scene change detection, where new observations are compared with an earlier record of the same environment. The interval between reference capture and subsequent queries allows advance reconstruction, which existing 3DGS-based methods use to render viewpoint-aligned references. To reduce the cost of online iterative optimization, we further use the reference scene for closed-form Bayesian updates and automatic-edit supervision of pixel prediction and output refinement. More details are provided in §1, 'Introduction'.

Expanded evaluation (2xdA: W2/Q2; MkVo: W2/W6/Q2). Keeping the original hyperparameter settings fixed, we extended the evaluation from PASLCD and CCS3D to CL-Splats and 3DGS-CD. Accuracy comparisons, progressive component ablations, and offline preparation and online runtime now cover all four datasets in Tables 1, 2, and 3 of the revised manuscript, respectively. The tables below summarize the main accuracy and online speed comparisons with O-SCD.

Accuracy comparison (mIoU ↑, %). O-SCD(O): online; O-SCD(R): offline-refined; Ours: our complete method, including output refinement.

MethodPASLCDCCS3DCL-Splats3DGS-CD
O-SCD(O)49.2449.9461.5767.14
O-SCD(R)55.9943.5956.8979.60
Ours67.0873.7771.1288.09

Complete online speed on RTX 5090 (FPS ↑). O-SCD(R) is offline-refined and has no corresponding online FPS.

MethodPASLCDCCS3DCL-Splats3DGS-CD
O-SCD(O)14.44.64.66.2
O-SCD(R)————
Ours17.514.813.911.8

Our complete method outperforms online O-SCD in accuracy on all four benchmarks. In the new runtime evaluation, it achieves 11.8–17.5 FPS, running 1.2–3.2× faster across datasets. More details are provided in Table 1 and Table 3B.

Component contributions and practical trade-offs (2xdA: W1/Q3; MkVo: W2/W5/Q1/Q2; 22dV: W1/Q2). To address reviewers' questions about component contributions and preparation costs, we compare three configurations offering different balances of accuracy, offline preparation, and online speed:

  • Base retains only the components required for our Bayesian updates (Rt0R_t^0, aggregated support StS_t, and historical support HtH_t), requiring no reference-edit supervision.
  • Standard adds reference-edit-trained suppression, LLR mapping, and the pixel prior, corresponding to Ours (pixel) in the original manuscript.
  • Full further uses reference-edit supervision to guide output refinement, corresponding to Ours (full) in the original manuscript and Ours in the tables above.

Configuration trade-offs on PASLCD (RTX 5090).

ConfigurationPreparation (s)mIoU (%)FPS
Base27.957.4027.0
Standard184.760.1327.7
Full215.867.0817.5

Preparation times are per scene instance, given an existing reference reconstruction. Base and Standard preparation times are estimates derived from the measured Full total, retaining preprocessing and remaining overhead.

Base offers competitive accuracy with minimal preparation; Standard generally improves accuracy through additional offline learning at similar online speed; Full adds offline contour fitting and online refinement for further accuracy gains. Similar trade-offs are observed on the other three datasets, with Base exceeding online O-SCD in accuracy and speed on all four benchmarks and Full adding 7.3–18.7 mIoU points over Base. Complete accuracy and runtime results are provided in Tables 1 and 3 of the revised manuscript.

AuthorsOpenReview ↗

Response to All Reviewers (2/2): Component Contributions and O-SCD Comparisons

Any contributions beyond the fitted prior and refinement? (2xdA: W1/Q3)

Yes. Our Bayesian framework alone improves both AP and mIoU, providing a strong foundation for further gains from reference-edit supervision.

We realised that the original ablations did not clearly explain component interactions, making the fitted prior appear to account for most of the gain. The reviewers' comments prompted progressive ablations, clarifying the contribution of Bayesian fusion and the further gains from reference-edit supervision. The original component-removal ablations have been moved to the appendix; the main text now presents progressive ablations across all four datasets.

PASLCD results are shown below (AP measures how well prediction scores rank changed pixels ahead of unchanged pixels.).

Configuration (cumulative)AP (%)mIoU (%)
Initial response R₀75.5252.43
+ Aggregated support Sₜ79.0753.62
+ Historical support Hₜ (Base)81.8457.40
+ LLR mapping82.5640.39
+ Response suppression83.8843.90
+ Fitted prior (Standard)83.8860.13
+ Contour filter85.1662.12
+ Object completion (Full)86.9267.08

All configurations preceding the fitted-prior row use a neutral prior.

  • Base: Bayesian fusion adds 6.3 AP points and 5.0 mIoU points over the initial response on PASLCD.
  • Standard: LLR mapping and suppression improve AP by 2.0 points but change the evidence scale, making the neutral prior unsuitable and reducing mIoU by 13.5 points. The fitted prior adjusts the decision boundary to this scale, recovering 16.2 mIoU points without changing the score ranking or AP. This large recovery therefore reflects the interaction between the learned responses and the prior. Together, these components improve mIoU over Base by 2.7 points.
  • Full: Contour filtering and object completion add 2.0 and 5.0 mIoU points, respectively, bringing the total gain over Base to 9.7 points.

A similar pattern appears on the other three datasets. More details are provided in Table 2 and Appendix G.4.

Do threshold fitting (prior) and output refinement improve O-SCD? (2xdA: Q3; 22dV: Q1)

Not consistently: some settings improve accuracy, while others reduce it, and performance remains well below our Full configuration.

Threshold fitting changes PASLCD mIoU by less than 0.1 points and improves it on the other three datasets, while refinement, alone or combined with threshold fitting, gives mixed results. All remain below our Full configuration. We evaluated threshold fitting and output refinement on O-SCD across all four datasets, fitting each contour model on its backend's reference-edit proposals. The table compares these transfers with our Standard and Full configurations in mIoU (%). The O-SCD baseline is quoted from Table 1 of the revised manuscript; the added-stage results come from separate evaluations.

MethodPASLCDCCS3DCL-Splats3DGS-CD
O-SCD(O)49.2449.9461.5767.14
O-SCD(O) + threshold fitting49.3355.3164.2368.34
O-SCD(O) + our refinement49.0262.5256.2276.65
O-SCD(O) + threshold fitting + our refinement49.2363.5552.8275.36
Standard60.1366.6767.3277.41
Full67.0873.7771.1288.09

With the same threshold-fitting and refinement procedures, Full exceeds O-SCD by 10.2–18.3 mIoU points across all four datasets. These results support the contribution of our Bayesian framework and the further benefits of reference-edit learning. More details are provided in Appendix G.5, "O-SCD Ablations and Output-stage Transfers".

AuthorsOpenReview ↗

Follow-up on our responses

Dear reviewers, thank you again for your time and feedback. We have added experiments and clarifications addressing your concerns in our responses and revised manuscript. As the discussion period draws to a close, we would appreciate your assessment of whether these additions resolve your main concerns, and would be happy to clarify any remaining points.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.