- 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 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 and the log-likelihood ratio mapper 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 , and spatial neighborhood covariance is approximated with a scalar parameter . 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
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. | Pose | Render | R0 | Infer. | Refine | Total | FPS | mIoU |
|---|---|---|---|---|---|---|---|---|
| O-SCD online | 25.2 | 0.6 | 5.8 | 37.7 | — | 69.3 | 14.4 | 49.24 |
| Base | 26.1 | 0.6 | 5.8 | 4.4 | — | 37.0 | 27.0 | 57.40 |
| Standard | 25.1 | 0.6 | 5.6 | 4.8 | — | 36.1 | 27.7 | 60.13 |
| Full | 25.2 | 0.6 | 5.6 | 5.3 | 20.3 | 57.2 | 17.5 | 67.08 |
CCS3D
| Config. | Pose | Render | R0 | Infer. | Refine | Total | FPS | mIoU |
|---|---|---|---|---|---|---|---|---|
| O-SCD online | 31.0 | 2.3 | 5.7 | 177.0 | — | 216.0 | 4.6 | 49.94 |
| Base | 30.9 | 2.1 | 5.9 | 12.5 | — | 51.6 | 19.4 | 55.05 |
| Standard | 30.3 | 2.0 | 5.6 | 13.1 | — | 51.0 | 19.6 | 66.67 |
| Full | 31.0 | 2.0 | 5.6 | 13.0 | 15.8 | 67.5 | 14.8 | 73.77 |
CL-Splats
| Config. | Pose | Render | R0 | Infer. | Refine | Total | FPS | mIoU |
|---|---|---|---|---|---|---|---|---|
| O-SCD online | 33.8 | 1.8 | 5.7 | 174.8 | — | 216.2 | 4.6 | 61.57 |
| Base | 34.0 | 1.6 | 5.7 | 12.3 | — | 53.7 | 18.6 | 63.85 |
| Standard | 33.4 | 1.7 | 5.7 | 12.5 | — | 53.3 | 18.8 | 67.32 |
| Full | 34.3 | 1.7 | 5.7 | 12.7 | 17.4 | 71.8 | 13.9 | 71.12 |
3DGS-CD
| Config. | Pose | Render | R0 | Infer. | Refine | Total | FPS | mIoU |
|---|---|---|---|---|---|---|---|---|
| O-SCD online | 53.8 | 1.4 | 6.1 | 100.5 | — | 161.9 | 6.2 | 67.14 |
| Base | 49.5 | 1.3 | 6.0 | 9.0 | — | 65.8 | 15.2 | 77.75 |
| Standard | 48.7 | 1.3 | 5.9 | 9.5 | — | 65.4 | 15.3 | 77.41 |
| Full | 49.3 | 1.3 | 5.9 | 9.7 | 18.8 | 85.0 | 11.8 | 88.09 |
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 frame | Frames | Median (ms) | P95 (ms) |
|---|---|---|---|
| 0–16 | 75 | 15.0 | 30.8 |
| 17–32 | 146 | 16.2 | 18.4 |
| 33–64 | 128 | 18.5 | 35.9 |
| 65–128 | 98 | 22.7 | 30.4 |
| 129–256 | 41 | 29.3 | 42.7 |
| >256 | 12 | 61.6 | 108.6 |
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".
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 adjusts the conditional mean and variance of each local Bayes factor (Eq. (12)), and setting 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 improves Full mIoU by approximately 0.07 points; the other 11 settings reduce it.
Both rules use the retention factor , where counts effective observations of a Gaussian. Discounting scales previous evidence by ; EMA additionally scales incoming evidence by . We evaluate both rules at 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.
| Update | Half-life H | Retention λ | mIoU ↑ | FPR ↓ | Recall ↑ |
|---|---|---|---|---|---|
| Original update (Table 1) | — | 1.0000 | 67.08 | 0.734 | 77.97 |
| Discounting | 1 | 0.5000 | 65.98 | 0.621 | 75.44 |
| Discounting | 2 | 0.7071 | 66.45 | 0.639 | 76.28 |
| Discounting | 4 | 0.8409 | 66.86 | 0.672 | 77.08 |
| Discounting | 8 | 0.9170 | 66.95 | 0.711 | 77.57 |
| Discounting | 16 | 0.9576 | 67.01 | 0.723 | 77.79 |
| Discounting | 32 | 0.9786 | 67.15 | 0.726 | 77.98 |
| EMA | 1 | 0.5000 | 65.68 | 0.615 | 75.00 |
| EMA | 2 | 0.7071 | 65.54 | 0.620 | 74.99 |
| EMA | 4 | 0.8409 | 65.26 | 0.634 | 74.78 |
| EMA | 8 | 0.9170 | 65.03 | 0.623 | 74.35 |
| EMA | 16 | 0.9576 | 64.67 | 0.622 | 73.92 |
| EMA | 32 | 0.9786 | 64.32 | 0.617 | 73.53 |
More details are provided in Appendix G.6, "Original sequences".
Half-Normal scale and spatial co-response (Q5)
How is determined across scenes, and how sensitive are the results? 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 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, provides a natural unit-scale default for the non-negative response increase. We use this setting across datasets and assess its sensitivity below. Testing 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.
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:
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".