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

• 2D-feature detectors struggle with viewpoint shifts and occlusions, explicit 3D pipelines often require heavy optimization, and benchmark-specific training limits cross-environment generalization.These challenges constrain change detection across environments and observation trajectories.Source: §1 · p. 2
• Argos uses geometric foundation model representations and co-visibility-aware inter-session comparison to jointly predict change masks and scene geometry from two unposed RGB image collections.Co-visibility helps distinguish an object's removal from its mere occlusion.Source: §1 · p. 2
• Argos-CD combines two large-scale synthetic datasets with a real-world evaluation set, and joint training supports gains across paired-image and video benchmarks, including synthetic-to-real transfer.Synthetic-only change supervision surpasses learned baselines fine-tuned on real data in real-world evaluations.Source: §1 · p. 2
• Argos-SLAM integrates Argos with VGGT-SLAM for online reconstruction and change updates in change-aware 4D mapping.The contribution connects learned change predictions to online reconstruction of a change-aware point-cloud map.Source: §1 · p. 2

• Geometric foundation models such as VGGT and VGGT-Ω predict camera poses and dense geometry through feedforward inference, and their latent representations implicitly capture epipolar geometry.Argos uses GFM latent representations and tokens for change detection rather than relying only on reconstruction outputs.Source: §2 · p. 2
• Scene-change methods range from pairwise learned features to multi-view comparison, but learned approaches can struggle with cross-domain transfer and zero-shot approaches often require costly reconstruction or optimization.Argos instead compares sessions directly in frozen GFM latent space and uses joint training across diverse datasets to improve transfer.Source: §2 · p. 3
• Dynamic-environment mapping systems commonly detect changes through explicit map geometry or object associations; Argos-SLAM combines online reconstruction and learned RGB change detection in a change-aware point-cloud map.The system uses SLAM for spatio-temporal structure and learned RGB change detection to improve identification beyond geometry alone.Source: §2 · p. 3

• The inputs are two sets of unposed RGB images from visits that may differ in length, trajectory, viewpoint, and overlap.Source: §3 · p. 3
• The goal is a binary change mask for every image, labeling changed content in the visit where it is physically present.This includes pixels for a new object in one visit and for an earlier object absent from a co-visible frame in the other.Source: §3 · p. 3
• A change-aware point-cloud map for each visit removes geometry belonging only to the other visit while retaining shared background and that visit's specific geometry.For example, querying the first visit removes a replacement bowl and recovers the scene with the earlier cup.Source: §3 · p. 3

4.1 MOTIVATION

• VGGT-Ω features align corresponding 3D locations across views; Argos uses pose and session information to compare visits and interpret inconsistent content as change.The design targets the ambiguity of 2D appearance differences caused by viewpoint or occlusion.Source: §4.1 · p. 4

4.2 MODEL ARCHITECTURE

• Argos jointly predicts change masks and scene geometry using a Camera Reader, inter-session cross-attention, and a DPT Fusion Head; masks disentangle visit-specific maps.Camera tokens augment patch features, while cross-attention compares features from the two visits.Source: §4.2 · pp. 4–5, Figure 3 · p. 5

4.3 NEW DATASET

• 4× more visually distinct places on average and up to 2× longer maximum video duration: Argos-CD-HSSD exceeds VSCD and SceneDiff on these dataset measures.Argos-CD comprises two synthetic datasets and a real-world evaluation set; its synthetic pipeline samples scene states, plans object- and area-covering traversals, and derives masks from object identities.Source: §4.3 · p. 5, Table 1 · p. 6

4.4 TRAINING

• Training combines SCD and video-SCD datasets, freezes VGGT-Ω components, and trains the three comparator modules on sequences from 2 to 64 frames.Evaluation uses held-out benchmark test sets and Argos-CD-real for additional out-of-distribution assessment.Source: §4.4 · p. 6

4.5 ONLINE ROBOT MAPPING SYSTEM

• Argos-SLAM uses Argos for submap geometry and cross-visit change detection, retrieving up to 64 relevant images for sub-second inference while asynchronous jobs update the map.Each completed job incorporates detected changes into the map accumulated through the current submap.Source: §4.5 · p. 6

5.1 EXPERIMENTAL SETUP

• The evaluation compares learned and zero-shot SCD and VSCD methods, four Argos training variants, and reports static IoU, change IoU, mIoU, and change-class macro F1.The variants are dataset-specific, synthetic-only, synthetic then real-world fine-tuned, and trained on all available datasets.Source: §5.1 · pp. 6–7

5.2 REAL-WORLD GENERALIZATION

• On SceneDiff, Ours-synthetic-only scores 39.84 change IoU versus 20.70 for the best fine-tuned learned baseline; on Argos-CD-real, it scores 20.87 versus 11.72.Ours-foundation matches or exceeds Ours-finetune; training VSCDNet with the same strategy yields 19.27 versus 62.54 change IoU on SceneDiff.Source: §5.2 · p. 7, Table 2 · p. 7

5.3 VIDEO SCENE CHANGE DETECTION (VSCD)

• On Argos-CD-SceneSmith, Argos increases change IoU by 42.01% and F1 by 27.91%, outperforming all baselines across VSCD benchmarks and metrics.Ours-foundation retains a similarly strong advantage.Source: Table 3 · p. 8, §5.3 · p. 8

5.4 SCENE CHANGE DETECTION (SCD) FOR PAIRED-IMAGE

• With one image per session, Argos achieves the highest change IoU and F1 on PSCD, ChangeSim, and VL-CMU-CD.The reported advantage therefore also applies to paired-image detection, not only video with additional context.Source: §5.4 · p. 8, Table 4 · p. 8

5.5 ABLATION STUDIES

• The Change Detection Head on GFM features already beats baselines; on VSCD, Camera Reader adds gains with Inter-session Cross-Attention but hurts in isolation.Replacing the GFM backbone with SAM degrades performance; the table caption also reports that Inter-session Cross-Attention and Camera Reader complement each other.Source: §5.5 · p. 8, Table 5 · p. 9

5.6 REAL-WORLD EXPERIMENT

• With inputs capped at 64 images, a change-detection pass takes 0.59 s; online submaps average 8.2 s and asynchronous change detection takes 6.9 s.Change updates typically finish before the next submap is produced.Source: §5.6 · p. 9

6 CONCLUSION

• Argos combines implicit 3D knowledge from GFMs with a co-visibility-aware spatio-temporal comparator for joint scene-change detection and 3D reconstruction.Training across diverse datasets, including the new benchmark, supports generalization across learned and zero-shot settings.Source: §6 · p. 9
• Argos-SLAM performs online change-aware 4D mapping over long sequences, bridging change detection and long-term robot mapping.Source: §6 · p. 10
• The results demonstrate the potential of implicit geometric knowledge for change detection and encourage further research on other spatial reasoning tasks.Source: §6 · p. 10

A.1 MODEL IMPLEMENTATION DETAILS

• The Camera Reader maps patch-token width from 2d to d/2 and back to d through two linear layers with GELU.The stated dimensions are 4096 → 1024 → 2048.Source: §A.1 · p. 14
• Cross-attention lets every token in one visit attend to all tokens in every frame of the other visit, in both directions.The camera and register tokens pass through unchanged; only patch tokens are updated.Source: §A.1 · p. 14
• Weighted binary cross-entropy is computed only on valid pixels, with change pixels up-weighted by 5 and unlabeled frames excluded.Frames without pixel-aligned labels receive an empty valid-pixel set.Source: §A.1 · p. 14
• Training interleaves paired-image and video streams by deficit scheduling so both data types receive balanced supervision and finish together.The training recipe also randomly swaps visit order with probability 0.5.Source: §A.1 · p. 14
• The Camera Reader and cross-attention modules are specified as randomly initialized, while the change head is warm-started from the pretrained depth head.Only the depth head’s final output layers are re-initialized for the change head.Source: §A.1 · p. 14

A.2 DATASET GENERATION

• Synthetic scenes generate three states by randomly removing movable objects while preserving support dependencies and retaining navigation-critical structures.Dependent objects are removed with their supports to maintain physical plausibility.Source: §A.2 · p. 14
• Coverage planning selects viewpoints to jointly cover scene objects and navigable areas, then connects them into an open-path TSP trajectory for each state.The trajectory-generation figure describes maximizing visible-object coverage with relatively few viewpoints.Source: §A.2 · p. 14, Figure 6 · p. 15
• Each state is rendered at 512 × 512 and 30 fps, with camera orientation varied independently of travel direction.This increases viewpoint diversity.Source: §A.2 · p. 14
• Change masks mark objects present in one state but absent from another, using object identities rather than pixel differences to produce exact per-frame labels.The changed objects receive a dedicated semantic label.Source: §A.2 · pp. 14–15
• Frames from independently generated traversals are matched across states using scene overlap and pose similarity, retaining the best cross-visit correspondences.Argos-CD-HSSD uses HSSD scenes in Habitat, while Argos-CD-SceneSmith uses procedural scenes and provides whole-house and per-room variants.Source: §A.2 · p. 15
• Argos-CD-real uses independently captured indoor revisits, with annotator point prompts propagated by SAM2 into dense masks for additions, removals, modifications, and deformable changes.The dataset comparison reports places, visit lengths, cross-visit pairs, and frames with ground-truth masks.Source: §A.2 · pp. 15–16

A.3 ONLINE SYSTEM IMPLEMENTATION DETAILS

• Each current-visit submap triggers a bounded change-detection job that processes keyframes with selected prior keyframes in one Argos pass and transfers predictions to sparse 3D points.The prior visit is retained while the current visit is reconstructed online with VGGT-SLAM.Source: §A.3 · p. 16
• Cross-visit pairing starts from verified loop closures, then uses descriptor matches and geometric reprojection, centroid-neighbor search, and a short temporal fallback for unmatched frames.Frames without a correspondence are excluded from change inference and cross-visit fusion.Source: §A.3 · p. 16
• Matching and inference run on a background worker to keep reconstruction responsive, with jobs constrained to Argos’s 64-image limit.The current-visit window is reduced as needed while keeping matched pairs together.Source: §A.3 · p. 16
• Pixel change probabilities and pairing confidences transfer to sparse 3D points, then observations within each world-space voxel are fused by confidence-weighted mean.The system thresholds fused probabilities to remove changed points and form a change-aware map.Source: §A.3 · p. 16

A.4 EVALUATION PROTOCOL

• For video benchmarks, paired-image methods receive SALAD-matched cross-visit frames; protocols sample 32 frames per VSCD visit, 16 for SceneDiff and Argos-CD-real, and 16 within 48-frame windows for longer synthetic sequences.ChangeSim excludes the removed class, while VSCD uses masks for both visits.Source: §A.4 · pp. 16–17

A.5 ADDITIONAL EXPERIMENTS

• With 128 reference frames on the 12 VSCD scenes O-SCD can register, Argos outperforms O-SCD, which exceeds VSCDNet in this comparison.O-SCD often fails registration under the original 32-reference/32-query setup.Source: Table 7 · p. 17, §A.5 · p. 18
• O-SCD predictions can be degraded by poor render quality or incorrect camera registration, including registration errors after large scene changes.The figure analysis notes these issues can occur even when reconstruction methods succeed.Source: §A.5 · p. 18, Figure 9 · p. 18
• Change IoU rises from 34.20 to 51.75 and F1 from 46.01 to 65.13 over TERDNet on ChangeSim under its original protocol.Argos also outperforms prior methods on the original VSCD benchmark across synthetic and real-world data and reported strata.Source: §A.5 · p. 18, Table 8 · p. 19
• On the original VSCD benchmark, Argos outperforms prior methods across synthetic and real-world data and variations in video length, graphics quality, and changed-object count.Table 9 reports frame-wise F1 by dataset and those strata.Source: §A.5 · p. 19, Table 9 · p. 19
• Runtime measurements time one batch-size-1 forward pass on an RTX 5090 as input images grow; SceneDiff runs out of 125 GB host memory at 128 images.O-SCD runtime is dominated by offline reconstruction and fails at eight or fewer images because COLMAP cannot register a usable reconstruction.Source: §A.5 · p. 19, Table 10 · p. 19

A.6 ADDITIONAL REAL-WORLD QUALITATIVE RESULTS

• Qualitative mapping examples cover online robot runs in workshop and common-space environments, plus offline iPhone-video mapping in an apartment and office.The figures include small- and medium-sized settings.Source: §A.6 · p. 19, Figure 10 · p. 20, Figure 11 · p. 20

A.7 ADDITIONAL QUALITATIVE RESULTS FROM CHANGE DETECTION DATASETS

• For SCD examples with labels on only one visit, Argos produces predictions with less noise; further qualitative figures cover SceneDiff and Argos-CD datasets.The additional figures include VSCD, Argos-CD-real, HSSD, and SceneSmith examples.Source: §A.7 · p. 20, Figure 14 · p. 21, Figure 15 · p. 22