acceptodds
Under review as a conference paper at ICLR 2027

Argos: Adapt Rich Geometric Priors for Generalizable Online Scene-Change-Detection

Abstract

Robots operating in dynamic environments require reliable detection of how their surroundings change over time. Existing learning-based methods largely rely on pairwise 2D image features, which struggle under large viewpoint changes and occlusions, are sensitive to noise, and show limited generalization across domains, while explicit 3D approaches typically require costly offline optimization. We show that the implicit 3D knowledge of Geometric Foundation Models (GFMs) provides a strong basis for addressing these limitations. We introduce Argos, which adapts GFM features for joint scene change detection and 3D reconstruction. To address data scarcity and take a step toward a foundation model for scene change detection, we introduce a large-scale benchmark comprising two synthetic datasets and one real-world dataset, and train jointly across diverse datasets to improve cross-domain generalization. We further introduce Argos-SLAM, a real- time system designed for robotics, which performs online change detection and change-aware 4D mapping. Across benchmarks, our framework substantially outperforms existing baselines, with gains of up to 42.01% in change IoU and 27.91% in F1, while supporting scalable deployment in changing real-world environments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Geometric Foundation Model, Change Detection, Robot Mapping

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_fCsuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Argos, a scene change detection framework that leverages geometric priors from a pretrained geometric foundation model. It introduces a camera-aware cross-session comparator for change detection, together with new datasets and an online SLAM system for change-aware 4D mapping. Experiments demonstrate strong performance on multiple benchmarks and promising synthetic-to-real generalization.

Critical strengths
  1. The idea of leveraging implicit 3D priors for scene change detection is well motivated. The proposed design is relatively simple yet achieves substantial improvements over existing methods.

  2. The experimental evaluation is extensive, including cross-domain generalization and real-world robotic deployment. The integration with SLAM also makes the work practically relevant.

Critical weaknesses
  1. The robustness to large-scale environmental changes remains unclear. In particular, significant illumination changes or substantial scene rearrangements may affect both feature matching and change detection. While the paper demonstrates promising cross-domain generalization, these challenging conditions are not explicitly evaluated. If suitable evaluation are unavailable, the authors should at least discuss these potential limitations and failure cases.

  2. The online system still relies on historical keyframe retrieval and geometric verification to establish cross-visit correspondences. This raises concerns about its robustness when the scene changes substantially, as relevant historical frames may fail to be matched and thus be excluded from change detection. It would be helpful to clarify how much the overall performance depends on this retrieval pipeline and discuss its potential failure modes.

Questions for authors

See the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to reviewer [1/2]

We thank the reviewer for the thoughtful comments and for recognizing the promising cross-domain generalization of our method. Both questions concern robustness under challenging conditions, and we address them with new analyses of illumination changes, scene rearrangement, and the retrieval pipeline, together with an explicit discussion of the remaining limitations. Below we indicate where the paper changed.

W1. The robustness to large-scale environmental changes remains unclear. In particular, significant illumination changes or substantial scene rearrangements may affect both feature matching and change detection. While the paper demonstrates promising cross-domain generalization, these challenging conditions are not explicitly evaluated. If suitable evaluation are unavailable, the authors should at least discuss these potential limitations and failure cases.

We agree that these conditions deserve explicit evaluation. We added a failure and limitation analysis to Sec. 6, with details in App. A.8.

Robust to moderate illumination changes. We simulate lighting changes by relighting every frame of the current visit while keeping the prior visit unchanged (Table 13). Mild changes barely affect the model, and even extreme darkening reduces change IoU by at most 2.7%. Extreme brightening causes the largest drop, because image saturation destroys image detail. Results of Ours-foundation, with relative change (%) to the unaltered input in parentheses:

LightingVSCD ChangeVSCD F1Argos-CD-real ChangeArgos-CD-real F1
None68.5153.3347.7441.24
Mild dark68.30 (-0.3)53.20 (-0.2)47.90 (+0.3)41.13 (-0.3)
Extreme dark66.66 (-2.7)52.45 (-1.7)46.96 (-1.6)40.10 (-2.8)
Mild bright67.49 (-1.5)53.01 (-0.6)47.19 (-1.2)40.95 (-0.7)
Extreme bright64.39 (-6.0)51.76 (-2.9)44.06 (-7.7)37.82 (-8.3)

Extreme real-world illumination. Beyond simulation, we newly evaluated the model qualitatively on real-world scenes with extreme lighting changes. In the most extreme case, an outdoor scene captured at noon and at night, the pretrained VGGT-Ω\Omega backbone fails to register the two visits, which leads to poor change detection (Fig. 6a). This failure stems mainly from the backbone's registration under extreme lighting rather than from our change-detection design, and our model can directly benefit as GFMs become more robust.

Substantial scene rearrangement. When furniture is heavily rearranged, VGGT-SLAM drifts (App. A.8, Fig. 13). Even under this drift, our pipeline still retrieves correct frames in parts of the scene and detects sensible changes in better-aligned areas, such as the kitchen and the table, which shows some robustness to SLAM drift. It does miss the largest changes, which we acknowledge as a limitation in App. A.8.

Limitations. Sec. 6 now states that our datasets mainly cover indoor, object-level changes, so very large-scale settings such as outdoor scenes are beyond the current scope. It also notes that small objects far from the camera in cluttered scenes are hard to detect (Fig. 6c).

AuthorsOpenReview ↗

Response to reviewer [2/2]

W2. The online system still relies on historical keyframe retrieval and geometric verification to establish cross-visit correspondences. This raises concerns about its robustness when the scene changes substantially, as relevant historical frames may fail to be matched and thus be excluded from change detection. It would be helpful to clarify how much the overall performance depends on this retrieval pipeline and discuss its potential failure modes.

We added this analysis to App. A.7.

Retrieval works well when SLAM is accurate. On the real-world sequences, fewer than 2% of keyframes are skipped in four of the five sequences, and 6.2% on the largest one, processed at real-time rate (Table 12). Retrieval is therefore reliable in typical robot deployment. Occasional skips are also tolerable: Argos-SLAM fuses change predictions into the 3D map, so detections from neighboring keyframes can still propagate change labels to the points of a skipped keyframe.

SequenceCaptureKeyframes (prior / current)Skipped
OfficeHandheld55 / 211.3%
HomeHandheld103 / 830%
AtriumRobot283 / 3240.3%
WorkshopRobot180 / 1590%
Full-floor loopRobot (real-time)841 / 8056.2%

Pose quality matters more than retrieval itself. Skips mostly come from SLAM drift. On Argos-CD-SceneSmith, skips concentrate in House scenes, whose long, fast trajectories were designed for data generation rather than SLAM. Each follows a single open path without revisits (Fig. 8), so VGGT-SLAM cannot close loops to correct the accumulated drift. As a result, change IoU is 76.56 on Room scenes but 49.53 on House scenes (Table 11), and accuracy drops as the skip rate grows (Fig. 12).

Failure mode. Under substantial scene changes, drift makes retrieval noisy and degrades change detection (Fig. 13). We acknowledge this dependence on SLAM as a limitation in Sec. 6, and reducing it is a promising direction for future work.

Thanks for the responses. I'll maintain my positive rating.

Reviewer_qLa1OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper works on the scene change detection (SCD) task, where the models are required to identify and segment the objects that appear or disappear in the scene without flagging false positives due to occlusion and viewpoint changes. The approach is to utilize pretrained geometric foundation models like VGGT-Ω\Omega, which show a strong attention response when objects are inconsistent between the views of both visits. To further extract the attention response, the authors propose to use Inter-session Cross-Attention, allowing tokens to use only views from the other visit in cross-attention. Camera Reader and DPT-Fusion Head are also proposed to further improve performance. The paper also proposes synthetic datasets Argos-CD-HSSD and Argos-CD-SceneSmith, a real-world dataset Argos-CD-real, and a SLAM model Argos-SLAM.

Critical strengths
  • Superior performance to previous work, with a clean and insightful observation of the attention maps of the pretrained VGGT-Ω\Omega backbone. The supplemental video and qualitative results show impressive detection accuracy of the proposed model, and the quantitative results surpass previous work by a non-negligible margin.
  • Even the synthetic-only trained model (ours-synthetic) shows good performance on the real dataset, demonstrating robustness to evaluation datasets.
  • Multifaceted contributions include the change-detection model, the benchmark, and the SLAM model.
Critical weaknesses
  • It is unclear what CDHead in Table 5 (Ablation) refers to. The term "Change Detection Head" is not defined in Section 4, and the closest guess may be the DPT Fusion Head or part of it.
  • There are only 8 scenes in the proposed Argos-CD-real dataset per Table 1, much smaller than the other ones, and therefore the contribution of the real dataset is weaker.
  • No quantitative evaluations are done for the proposed Argos-SLAM model other than the runtime comparisons (see Section 5.6)
  • In Table 5, it seems that most of the performance gain is contributed by the stronger backbone and the unclear CDHead, while the contribution of the architecture novelty (CrossAttention, CameraReader) is marginal.
  • Regarding the previous weakness, it may be argued that considering the approach of using a pretrained 3D backbone is also part of the paper's novelty; however, this still raises a concern about whether the approach works with general Geometric Foundation Models, as changing the backbone shows poor results in Table 5. I would suggest the authors show attention maps of other backbones in rebuttal to demonstrate the generality of their observation.
Questions for authors
  • Are v0 and v1 standard terms in the field? They are clear after being defined in Section 3, but in Figure 1 it might be clearer if the authors used "prior visit" and "current visit" or similar terms instead.
  • I appreciate the authors stating that the code and dataset will be released in the reproducibility statement. Will the code for generating the synthetic data and the model weights also be released?
  • In Table 2, there was a near -8% difference in IoU of the changed objects between "ours-foundation" and "ours-finetune", and it appears that including more datasets in the training process hurts the performance on the Argos-CD-real. Can the authors explain why this would happen?
4 comments
AuthorsOpenReview ↗

Response to reviewer [1/3]

We thank the reviewer for the careful reading and for pointing out several points that requires further clarification. We have clarified each of them to improve the presentation and added new experiments where requested. Below we indicate where the paper changed.

W1. It is unclear what CDHead in Table 5 (Ablation) refers to. The term "Change Detection Head" is not defined in Section 4, and the closest guess may be the DPT Fusion Head or part of it.

We apologize for the confusion. CDHead is the DPT Fusion Head (Sec. 4.2). We renamed it DPTHead in Table 5 and made the naming consistent throughout the paper.

W2. There are only 8 scenes in the proposed Argos-CD-real dataset per Table 1, much smaller than the other ones, and therefore the contribution of the real dataset is weaker.

We acknowledge that 8 scenes is small. We have strengthened the real-world dataset by collecting more data, and the new data also covers challenging scenarios that prior real-world datasets lack.

  • More data. We collected additional data, and Argos-CD-real now has 19 scenes with 12,367 annotated views (Table 1).
  • New, challenging scenarios. We deliberately captured scenes such as cluttered workshops with tiny changes and revisits traversed in opposite directions (Sec. 4.3). As a result, learned baselines reach at most 13.50 change IoU on Argos-CD-real, compared with 20.70 on the SceneDiff dataset (Table 2). This added difficulty is itself a contribution to the field.

All methods are re-evaluated on the expanded 19-scene Argos-CD-real (Table 2; †^\dagger zero-shot). The SceneDiff change IoU is shown for comparison:

MethodSceneDiff ChangeArgos-CD-real StaticArgos-CD-real ChangeArgos-CD-real mIoUArgos-CD-real F1
CSCDNet-finetune15.2592.6110.6451.6313.60
C-3PO-finetune17.8892.266.9649.6110.86
GeSCF †^\dagger5.0463.616.1534.8813.50
RobustSCD-finetune18.9791.7813.5052.6416.55
ZSSCD †^\dagger6.7477.533.1240.338.27
TERDNet-finetune20.3291.7811.3851.5816.56
VSCDNet10.4887.229.0548.1413.39
VSCDNet-finetune20.7093.718.9751.3414.61
VSCDNet-foundation19.2790.768.6849.7216.51
SceneDiff †^\dagger30.3195.5532.9264.2335.52
Ours-synthetic-only39.8494.7431.1862.9633.95
Ours-finetune59.9197.1350.9274.0341.51
Ours-foundation62.5497.0847.7472.4141.24

On the expanded set, our model remains the best by a large margin: Ours-finetune reaches 50.92 change IoU, compared with 32.92 for the strongest baseline, the zero-shot SceneDiff.

  • Extensible. The dataset is annotated with the SceneDiff tool, which we modified to be more user-friendly and faster (App. A.3). We will release the modified tool so that users can collect and annotate more data in the same way.

W3. No quantitative evaluations are done for the proposed Argos-SLAM model other than the runtime comparisons (see Section 5.6)

We added a system-level quantitative evaluation (App. A.7), summarized in Sec. 5.6. We run the full system, retrieval included, on all 90 Argos-CD-SceneSmith test pairs, and count ghost points: stale geometry that matches the other visit's ground truth within 5 cm but not the current one. Results from Table 11:

ScopePairsChangemIoUF1Ghost, naiveGhost, oursRemoved
Room7876.5687.5765.0642.1K22.9K41.9%
House1249.5373.3344.6653.7K45.1K14.1%
All9068.8683.5159.0543.6K25.9K38.2%
  • Accuracy is preserved in the full system. Change detection with retrieval included is consistent with the per-pair benchmark.
  • Effective change-aware mapping. Compared with naively fusing both visits, Argos-SLAM removes 38.2% of ghost points, showing that our change detection keeps the map consistent with the current scene.

App. A.7 also analyzes how often keyframes are skipped and what this costs (Fig. 12, Table 12); please also see our response to Reviewer 1, Q1-Part 2.

AuthorsOpenReview ↗

Response to reviewer [2/3]

W4. In Table 5, it seems that most of the performance gain is contributed by the stronger backbone and the unclear CDHead, while the contribution of the architecture novelty (CrossAttention, CameraReader) is marginal.

We agree that the GFM backbone provides the largest gain, which is the core message of our paper: GFM features are a strong basis for change detection. Our architecture is intentionally simple yet effective. It adds no backbone-specific tricks, hand-crafted heuristics, or backbone-specific components on top of VGGT-Ω\Omega, so it transfers directly to other GFM backbones such as π3\pi^3, and even to 2D backbones such as DINOv2 (W5). At the same time, its components contribute substantially on VSCD (Table 5):

BackboneDPTHeadCrossAttnCamReaderVSCD mIoUVSCD F1PSCD mIoUPSCD F1
DINOv2✓\checkmark✓\checkmark✓\checkmark73.4541.3274.3465.37
VGGT-Ω\Omega✓\checkmark72.1642.5279.2769.26
VGGT-Ω\Omega✓\checkmark✓\checkmark80.7050.0178.2969.08
VGGT-Ω\Omega✓\checkmark✓\checkmark✓\checkmark83.4653.1479.6869.89
  • Adding CrossAttention to VGGT-Ω\Omega + DPTHead improves VSCD mIoU by 8.5 and F1 by 7.5.
  • Adding CameraReader on top of CrossAttention brings a further 2.8 mIoU and 3.1 F1 on VSCD. CameraReader is designed to complement CrossAttention: it injects camera information that guides the cross-visit comparison, which is why it helps most when combined with CrossAttention.
  • The full model improves VSCD mIoU by 11.3 and F1 by 10.6 over VGGT-Ω\Omega + DPTHead. This is comparable to the gain from switching the backbone from DINOv2 to VGGT-Ω\Omega (+10.0 mIoU, +11.8 F1), so the architecture matters as much as the backbone choice.
  • On PSCD, which has small viewpoint changes, the components matter less, as expected for a design that targets cross-view comparison.

We also note that SAM is not a GFM. Table 5 now includes DINOv2 and π3\pi^3 for a fairer comparison (see W5).

W5. Regarding the previous weakness, it may be argued that considering the approach of using a pretrained 3D backbone is also part of the paper's novelty; however, this still raises a concern about whether the approach works with general Geometric Foundation Models, as changing the backbone shows poor results in Table 5. I would suggest the authors show attention maps of other backbones in rebuttal to demonstrate the generality of their observation.

Following this suggestion, we trained the full architecture with two more frozen backbones, the GFM π3\pi^3 and the 2D model DINOv2, using the same training recipe (Table 5, Sec. 5.5).

BackboneVSCD Chg.VSCD mIoUVSCD F1PSCD Chg.PSCD mIoUPSCD F1
SAM (2D)27.8959.9030.6446.0869.0758.59
DINOv2 (2D)50.0773.4541.3255.2974.3465.37
π3\pi^3 (GFM)64.3981.1850.8355.1174.1264.44
VGGT-Ω\Omega (GFM)68.6983.4653.1464.5679.6869.89
  • Our approach generalizes to another GFM. With π3\pi^3, the same architecture reaches 64.39 change IoU on VSCD, close to VGGT-Ω\Omega (68.69) and far above the 2D backbones. This confirms that our observation holds beyond a single GFM.
  • Attention maps. As suggested, App. A.1 (Fig. 7) visualizes the cross-visit attention of all four backbones. Both GFMs attend sharply to the queried object, and VGGT-Ω\Omega stays near zero where the object is absent, while the 2D backbones attend to unrelated regions. This further supports our motivation.

Q1. Are v0 and v1 standard terms in the field? They are clear after being defined in Section 3, but in Figure 1 it might be clearer if the authors used "prior visit" and "current visit" or similar terms instead.

There is no standard notation in the field yet. Following the suggestion, we added "before change: v0v_0" and "after change: v1v_1" labels to Fig. 1.

Q2. I appreciate the authors stating that the code and dataset will be released in the reproducibility statement. Will the code for generating the synthetic data and the model weights also be released?

Yes. We will release the data-generation code and model weights together with the code and datasets. App. A.3 describes the generation pipeline step by step.

AuthorsOpenReview ↗

Response to reviewer [3/3]

Q3. In Table 2, there was a near -8% difference in IoU of the changed objects between "ours-foundation" and "ours-finetune", and it appears that including more datasets in the training process hurts the performance on the Argos-CD-real. Can the authors explain why this would happen?

We apologize for the unclear description and clarify it here. Both variants are trained on the same datasets and differ only in the training schedule (Sec. 5.1, Model Variants). Ours-finetune is first trained on the synthetic video datasets and then fine-tuned on the real-world data, so it is optimized for the real-world data distribution. Ours-foundation is trained on all datasets jointly and must balance all scenarios. Results from Table 2 on the expanded 19-scene Argos-CD-real:

VariantTraining dataSceneDiff ChangeArgos-CD-real Change
Ours-synthetic-onlySynthetic39.8431.18
Ours-finetuneSynthetic, then real59.9150.92
Ours-foundationSynthetic + real, jointly62.5447.74
  • Small gap, broader model. On the expanded set, the gap on Argos-CD-real is only 3.2 change IoU, and Ours-foundation is better on the SceneDiff dataset. A single jointly trained model thus stays competitive across datasets.
  • More data helps. Ours-foundation, which adds the real-world data, clearly outperforms Ours-synthetic-only.

I appreciate the authors' clarification and the evidence provided through the additional experiments, and I would suggest the authors include the enlarged dataset (Table in W2) and ablation (Table in W4) in their final version. Meanwhile, I'm increasing my rating for this paper.

Reviewer_DpbFOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Argos, a feed-forward model for detecting scene changes between two visits to the same place. It keeps VGGT-Ω frozen and trains three small modules on top: a Camera Reader, an inter-session cross-attention, and a DPT-style head that outputs a change mask for every image. Because the original pose and depth heads still run, one forward pass gives change masks, camera poses and depth, which together produce change-aware 4D point-cloud maps.

Furthermore, the authors also contribute Argos-CD, two synthetic datasets plus a small real evaluation set, and plug the model into VGGT-SLAM 2.0 for online mapping. It beats prior methods by a large margin on both paired-image and video benchmarks.

Critical strengths
  1. The design is simple: a frozen geometric foundation model plus three lightweight modules. The ablations separate the effect of the backbone from the two added modules, and the four model variants make it clear what the training data contributes.
  2. The method copes with large viewpoint changes and produces object-shaped change masks without any instance segmentation.
  3. The paper is clearly written and easy to follow.
Critical weaknesses
  1. The sim-to-real claim is weak on the authors’ own real set. On Argos-CD-real, the synthetic-only model beats zero-shot SceneDiff by only 1.4 change IoU and is lower in F1. With just 8 pairs and no variance reported, this margin says little. Real-data fine-tuning adds about 20 IoU points on SceneDiff Dataset, meaning that real data is important. I would suggest authors modifying the claim.
  2. The only backbone comparison is SAM in this paper, which differs from VGGT-Ω in data, scale and training objective. There, this comparison is not very fair that does not isolate the effect of geometric priors. Results with other geometry models, such as VGGT, π3, MASt3R, or with a 2D backbone such as DINOv2, would make the argument much stronger. This is encouraged but not required.
Questions for authors
  1. The paper has no failure analysis in experiments. The proposed Camera Reader relies on the frozen backbone’s camera tokens, so how does the method behave when pose estimation degrades, such as, under large viewpoint gaps, lighting changes or many moving objects? In Argos-SLAM, frames with no retrieved match are skipped. How often does this happen, and what does it cost? The system results are qualitative and runtime only, with no 3D accuracy for the map.
  2. The output is a binary mask without object identity, so a moved object appears as an unrelated removal and addition. This is outside the paper’s scope, but instance-level association would make the system more useful for applications such as scene graph reconstruciton, and could be discussed as future work.
3 comments
AuthorsOpenReview ↗

Response to reviewer [1/3]

We sincerely thank the reviewer for the positive assessment and for recognizing our contributions. We addressed all questions, including the optional ones, because we agree that these additions make our argument stronger and the evaluation more holistic. Below we indicate where the paper changed.

W1. The sim-to-real claim is weak on the authors’ own real set. On Argos-CD-real, the synthetic-only model beats zero-shot SceneDiff by only 1.4 change IoU and is lower in F1. With just 8 pairs and no variance reported, this margin says little. Real-data fine-tuning adds about 20 IoU points on SceneDiff Dataset, meaning that real data is important. I would suggest authors modifying the claim.

We appreciate this careful observation and agree that the original wording was too strong. Sec. 1 now states "promising real-world generalization" instead of strong sim-to-real transfer.

Our inference is deterministic, so repeated evaluation yields identical results, and retraining with multiple seeds was not feasible. To obtain a more reliable estimate, we instead expanded Argos-CD-real from 8 to 19 scenes (Table 1). Results on the expanded set (from Table 2; †^\dagger zero-shot):

MethodSceneDiff ChangeSceneDiff F1Argos-CD-real ChangeArgos-CD-real F1
RobustSCD-finetune18.9721.2513.5016.55
VSCDNet-finetune20.7023.848.9714.61
SceneDiff †^\dagger30.3137.5832.9235.52
Ours-synthetic-only39.8437.4231.1833.95
Ours-finetune59.9144.7350.9241.51
Ours-foundation62.5445.9847.7441.24
  • Promising generalization without real data. Trained only on synthetic data, our model more than doubles the change IoU of the best fine-tuned learned baseline on Argos-CD-real (31.18 vs. 13.50), and it comes close to the zero-shot SceneDiff (32.92), which relies on costly per-scene optimization. Although we have made the claim more conservative, these results still show that our model transfers from synthetic to real data much better than other learned methods.
  • Real data further helps. We agree with the reviewer that real data is important. Fine-tuning raises change IoU to 50.92 on Argos-CD-real and 59.91 on the SceneDiff dataset, far beyond all baselines.

W2. The only backbone comparison is SAM in this paper, which differs from VGGT-Ω\Omega in data, scale and training objective. There, this comparison is not very fair that does not isolate the effect of geometric priors. Results with other geometry models, such as VGGT, π3\pi^3, MASt3R, or with a 2D backbone such as DINOv2, would make the argument much stronger. This is encouraged but not required.

Although this was optional, we ran the suggested experiments because they make our argument stronger. We added DINOv2 (2D, self-supervised) and π3\pi^3 (GFM) as backbones. All backbones are frozen and use the same full architecture and training recipe; only the backbone changes (Table 5, Sec. 5.5).

BackboneVSCD Chg.VSCD mIoUVSCD F1PSCD Chg.PSCD mIoUPSCD F1
SAM (2D)27.8959.9030.6446.0869.0758.59
DINOv2 (2D)50.0773.4541.3255.2974.3465.37
π3\pi^3 (GFM)64.3981.1850.8355.1174.1264.44
VGGT-Ω\Omega (GFM)68.6983.4653.1464.5679.6869.89
  • The geometric prior is the key. On VSCD, which has large viewpoint changes, both GFMs outperform the 2D backbones by a large margin (π3\pi^3 +14.3 and VGGT-Ω\Omega +18.6 change IoU over DINOv2). This isolates the benefit of geometric pretraining, as the reviewer suggested.
  • Consistent with our motivation. On PSCD, which has small viewpoint changes, DINOv2 and π3\pi^3 are on par, and VGGT-Ω\Omega remains best. The geometric prior helps most under large viewpoint changes; without them, GFM features contribute less, though VGGT-Ω\Omega still outperforms both 2D backbones.

App. A.1 (Fig. 7) further compares the cross-visit attention of all four backbones. Where the queried object is absent, GFM attention stays low, while the 2D backbones attend to unrelated regions. We thank the reviewer again for suggesting this insightful comparison.

AuthorsOpenReview ↗

Response to reviewer [2/3]

Q1-Part 1. The paper has no failure analysis in experiments. The proposed Camera Reader relies on the frozen backbone’s camera tokens, so how does the method behave when pose estimation degrades, such as, under large viewpoint gaps, lighting changes or many moving objects?

We agree that a failure analysis helps readers understand the limits of the model and encourages future work to address them. We added a failure analysis to Sec. 6 (Fig. 6), with details in App. A.8.

  • Robust to moderate lighting changes. We relight every frame of the current visit and keep the prior visit unchanged (App. A.8, Table 13). Mild changes barely affect the model, and even extreme darkening reduces change IoU by at most 2.7%. Extreme brightening causes the largest drop, because clipped highlights destroy image detail. Results of Ours-foundation, with relative change (%) to the unaltered input in parentheses:
LightingVSCD ChangeVSCD F1Argos-CD-real ChangeArgos-CD-real F1
None68.5153.3347.7441.24
Mild dark68.30 (-0.3)53.20 (-0.2)47.90 (+0.3)41.13 (-0.3)
Extreme dark66.66 (-2.7)52.45 (-1.7)46.96 (-1.6)40.10 (-2.8)
Mild bright67.49 (-1.5)53.01 (-0.6)47.19 (-1.2)40.95 (-0.7)
Extreme bright64.39 (-6.0)51.76 (-2.9)44.06 (-7.7)37.82 (-8.3)
  • Degraded pose under extreme lighting. Beyond simulation, we newly evaluated the model qualitatively on real-world scenes with extreme lighting changes. In the most extreme case, an outdoor scene captured at noon and at night, the pretrained VGGT-Ω\Omega backbone fails to register the two visits, which leads to poor change detection (Fig. 6a). This is a limitation of the backbone's pose estimation rather than of our change-detection design, and our model can directly benefit from more robust GFMs in the future.
  • Large scene rearrangement. When furniture is heavily rearranged, VGGT-SLAM drifts and retrieval becomes noisy. Even so, our model still detects changes in better-aligned areas, although it misses the largest changes (App. A.8, Fig. 13).
  • Moving objects. Our formulation compares two static scene states and does not explicitly model dynamic objects. Although an ideal change-aware 4D map should capture all kinds of changes, including dynamic objects, this work addresses only the quasi-static part. We now state this in the limitations (Sec. 6) and leave dynamic objects to future work.
  • Other cases. We also show semantic changes and tiny objects in cluttered scenes (Fig. 6b–c).

Q1-Part 2. In Argos-SLAM, frames with no retrieved match are skipped. How often does this happen, and what does it cost? The system results are qualitative and runtime only, with no 3D accuracy for the map.

We added a system-level evaluation of Argos-SLAM (App. A.7), summarized in Sec. 5.6.

Skips are rare in deployment. On the real-world sequences, fewer than 2% of keyframes are skipped in four of the five sequences, and 6.2% on the largest one, processed at real-time rate (Table 12). This shows that our pipeline is well suited to robot deployment.

SequenceCaptureKeyframes (prior / current)Skipped
OfficeHandheld55 / 211.3%
HomeHandheld103 / 830%
AtriumRobot283 / 3240.3%
WorkshopRobot180 / 1590%
Full-floor loopRobot (real-time)841 / 8056.2%

Cost of skipping and 3D map quality. We run the full system, retrieval included, on all 90 Argos-CD-SceneSmith test pairs, and conservatively score skipped keyframes as unchanged. Map accuracy is mainly determined by VGGT-SLAM's pose and depth, and changed geometry is a small part of a large scene, so standard reconstruction metrics barely reflect our change detection. We therefore count ghost points: points that match the other visit's ground truth within 5 cm but not the current one, i.e., stale geometry. Results from Table 11:

ScopePairsChangemIoUF1Ghost, naiveGhost, oursRemoved
Room7876.5687.5765.0642.1K22.9K41.9%
House1249.5373.3344.6653.7K45.1K14.1%
All9068.8683.5159.0543.6K25.9K38.2%
  • Accuracy is preserved in the full system. Change detection with retrieval included (83.51 mIoU, 59.05 F1) is consistent with the per-pair benchmark.
  • Cleaner maps. Compared with naively fusing both visits, Argos-SLAM removes 38.2% of ghost points overall and 41.9% on Room scenes.
  • House scenes are harder for SLAM. Their fast, sparse trajectories were designed for data generation rather than SLAM evaluation. Each follows a single open path without revisits (Fig. 8), so VGGT-SLAM cannot close loops to correct the accumulated drift, which causes more skipped keyframes. Fig. 12 shows that accuracy drops as the skip rate grows, as expected.
AuthorsOpenReview ↗

Response to reviewer [3/3]

Q2. The output is a binary mask without object identity, so a moved object appears as an unrelated removal and addition. This is outside the paper’s scope, but instance-level association would make the system more useful for applications such as scene graph reconstruciton, and could be discussed as future work.

We agree that instance-level association would make the system more useful, and we added it to the limitations in Sec. 6. Argos outputs binary masks without object identity, so it cannot natively detect moved objects. We keep the framework general on purpose: it detects change at whatever granularity the data provides, and in principle can detect non-object changes such as structural changes, for which we currently lack data. Building on a more compact representation such as a scene graph to maintain an instance-level change list is a promising direction for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comments

We thank all reviewers for their constructive feedback. The revised paper includes the following changes:

New experiments

  • Backbone study (R1, R2): DINOv2 and π3\pi^3 backbones (Table 5) and attention comparison (App. A.1). The results confirm that GFM features benefit multi-view change detection and that our observation generalizes beyond VGGT-Ω\Omega.
  • Argos-SLAM evaluation (R1, R2, R3): system-level change detection, ghost points, and keyframe skip analysis (App. A.7, summarized in Sec. 5.6). System-level change detection is consistent with the benchmark results, ghost-point removal shows effective change-aware mapping, and low skip rates in real-world trials show reliable frame retrieval.
  • Failure analysis (R1, R3): characterizes the limits of the model under illumination changes, scene rearrangement, semantic changes, and tiny objects (Sec. 6, App. A.8).
  • Larger real-world set (R1, R2): Argos-CD-real expanded from 8 to 19 scenes, with all methods re-evaluated (Tables 1–2).

Revisions

  • Claims (R1): more conservative real-world claim (Sec. 1).
  • Limitations and future work (R1, R3): a new paragraph acknowledges limitations beyond our current scope, such as dynamic objects and instance-level association, and suggests future directions (Sec. 6).
  • Clarity and release (R2): clearer naming (DPTHead) and notation (Fig. 1), annotation-tool details (App. A.3), and release of data-generation code and model weights.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.