acceptodds
Under review as a conference paper at ICLR 2027

CLASP: Complementary Chunk and Long Passes for Feed-Forward 3D Reconstruction

Abstract

Feed-forward 3D reconstruction faces two distinct failures as sequences grow: camera predictions can deteriorate in a full-sequence pass while overlapping chunks accumulate pose and scale errors through their alignment chain. We present CLASP (Chunk-Linked, Aligned-Scale Pose graph), a training-free framework that combines both inference contexts of the same frozen model in one pose graph. Chunks supply local relative motion and output geometry. The long pass supplies a common scale reference through dense depth ratios between predictions of the same frame, independently of camera motion and overlap alignment. It also contributes long-range pose constraints where they agree with a chunk-derived trajectory. Separating these roles retains useful scale information when long-pass pose constraints are rejected, allowing one formulation to handle long and short captures without a dataset-specific switch. Experiments with three backbones on six datasets show that CLASP lowers trajectory error on long driving sequences and keeps pose accuracy close to or above a full-sequence pass on shorter captures.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed-forward 3D Reconstruction, neural geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 5 reviews, above 85% of the others.
Reviewer_X5szOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

CLASP addresses complementary failure modes in feed-forward 3D reconstruction: full-sequence inference can produce unreliable camera trajectories, while overlapping chunks accumulate alignment and scale errors. It runs the same frozen reconstruction backbone on both contexts, estimates chunk-scale corrections from corresponding depth predictions, selectively admits long-pass relative poses, and combines these measurements with local links and loop closures in a robust pose graph. Reconstruction then reuses the rescaled chunk point maps.

The paper evaluates three backbones on 51 sequences from six datasets. Its central contribution is the observation that full-sequence depth can remain a useful scale reference even when full-sequence poses fail.

Critical strengths
  • Separating the reliability of depth-scale predictions from camera-pose predictions is a useful insight. Same-frame depth ratios provide a direct connection between each chunk and a shared reference, avoiding cumulative scale estimation through successive overlaps. This measurement design is the strongest original aspect of the paper.

  • Evaluation across three backbones and six datasets provides substantial breadth. The appendix contains meaningful diagnostics, scale-term ablations, threshold sweeps, and failure cases. In particular, Table 12 shows that removing long-pass depth constraints increases π³/4Seasons ATE from 93.51 m to 242.15 m, providing convincing evidence that these constraints contribute materially.

  • The overall division of responsibilities between the two inference contexts is easy to understand. Coordinate conventions and most solver settings are documented. The authors also disclose the use of cached predictions, additional computational cost, and cases where their assumptions fail.

  • Significance of contributions: Improving frozen geometric models without retraining is practically valuable. The improvements are substantial in several settings: with VGGT-Ω on KITTI-360, ATE decreases from 38.18 m for VGGT-Long to 20.26 m for CLASP. The observation that different outputs of one model can retain different levels of reliability deserves further investigation.

Critical weaknesses

W1.The reliability-selection mechanism is less established than the scale-correction mechanism.

Agreement with a chunk-derived trajectory cannot distinguish a correct long-pass constraint from a shared error, and can reject useful constraints when the chunk trajectory has drifted. The appendix acknowledges this limitation, but does not quantify its practical frequency. Furthermore, Eq. (7) retains per-frame long-pass rotation priors even when long-span pose edges are rejected; these priors influence relative rotations and do more than fix a global reference.

W2.The reported comparisons change several components simultaneously, including graph structure, loop handling, and access to full-sequence predictions.

A particularly informative missing control is VGGT-Long augmented with the proposed depth-scale measurements. Additionally, Appendix A.5 states that reported results reuse loop closures verified in an earlier full-pipeline run. Since scale correction affects loop verification, cached constraints may preserve information from components subsequently removed in an ablation. The reported agreement with fresh execution on π³/4Seasons is reassuring, but does not resolve this issue across configurations.

W3.The main ablation contains a concrete inconsistency.

Table 4 and Appendix A.5 describe the final step as restoring long-range translations, yet the 7-Scenes rotation error changes from 3.72° to 3.61°. Under Eqs. (7)–(8), changing translation constraints alone cannot alter the already recovered rotations. Either additional rotation factors or weights change, or the ablation description is inaccurate. This matters because Table 4 is the principal evidence separating the two long-pass roles.

W4.Missing related work or comparisons

The literature coverage is broad, but the distinction between the proposed measurement source and established graph machinery needs greater precision. VGGT-SLAM 2.0 is already cited, yet its use of keyframe-level graph variables and intra-submap pose constraints deserves a direct technical comparison. Its published formulation makes clear that moving beyond submap-only variables is not itself unique to CLASP. VGGT-SLAM 2.0, Section IV. The missing controlled depth-augmentation baseline is more consequential than adding a long list of additional methods.

Questions for authors
  1. What exactly changes between the last two rows of Table 4? Why does RRA change from 3.72° to 3.61° if only long-range translations are restored? Please specify all active factors and priors for each variant.

  2. How are cached loop constraints handled in the ablations? Were their acceptance decisions or measured translations obtained using the components being removed? Are measurements recomputed or rescaled after changing chunk scales?

  3. How much improvement remains when the depth-scale reference is added to VGGT-Long while retaining its graph and loop pipeline? This would isolate the paper’s central measurement contribution.

  4. Does span-level admission improve over admitting all candidate spans with the same robust weighting? Please report accepted-span counts and clarify the contribution of the persistent long-pass rotation priors.

  5. How are the ground-truth chunk scales in Tables 10 and 12 constructed? Please define the alignment and MAD calculation, particularly for driving datasets without the dense depth used in the geometry evaluation.

  6. Can the authors provide an exact evaluation manifest and configuration record? This should identify sequence/frame selections, baseline code versions, threshold-selection data, and the complete runtime and memory cost of the VGGT-Ω configuration used in the main tables.

5 comments
AuthorsOpenReview ↗

Response to Reviewer X5sz — Part 1

We thank the reviewer for the thorough review. We address the concerns below and are open to further discussion and questions. Several points are presented unclearly in the submission and we correct them in the order of the questions. Cells are AUC@30 ↑ / ATE (m) ↓, with π³ unless stated, and Long (44) is the mean over the 44 sequences of KITTI Odometry, KITTI-360, 4Seasons, TartanAir and Virtual KITTI.

Q1, W3: What changes between the last two rows of Table 4.

We clarify that the labels in Tab. 4 are misleading. As Sec. 3.3 states, admitted spans enter both the rotation solve (Eq. 7) and the position solve (Eq. 8). The table below adds the long-pass inputs one at a time on 7-Scenes, with the predictions and loop set of Table 4; rows 2 to 4 are the three graph rows of Table 4. Every row contains the chunk links and verified loops. The rotation prior enters in row 2, the long-pass scale in row 3 and the admitted spans in row 4:

VariantAUC@30 ↑ATE (m) ↓RRA (°) ↓RTA (°) ↓
Chunk graph, no long-pass input0.7180.0844.379.86
+ rotation prior (Table 4 "Chunk graph")0.7350.0813.729.41
+ long-pass scale0.7340.0813.729.37
+ admitted spans (full CLASP)0.7870.0583.617.42
CLASP without long-pass pose input (scale kept)0.7160.0834.379.83

Within a single room the chunks already share a consistent scale, so the depth-ratio measurement has little to correct here (row 3) and pays off on the driving data instead. The full pass, in contrast, stays accurate over long spans of a room-scale sequence, and the admission test lets these spans in, where they tie distant frames together in both the rotation and the position solve and remove the drift that the chunk links accumulate (AUC@30 from 0.734 to 0.787, row 4). The RRA change from 3.72° to 3.61° comes from the rotation part of these spans. With only their translation part, RRA stays at 3.72°. Without the rotation prior and the admitted spans, CLASP falls back to the first row (last row). So on 7-Scenes the gain comes from long-pass pose information, chiefly the spans that the admission test lets in.

Q2, W2: Cached loops in the ablations.

All rows of an ablation share one loop set, verified in an earlier run of the full pipeline. Its acceptance decisions therefore used the long-pass scale. To test whether the cached loops favour some rows, we ran the variant without any long-pass input (first row above) end to end, so that it verifies its own loops (AUC@30, cached loop set → own loops):

Datasetπ³VGGTVGGT-Ω
7-Scenes0.718 → 0.7190.716 → 0.7390.813 → 0.784
KITTI Odo.0.824 → 0.8160.611 → 0.5990.813 → 0.780
KITTI-3600.679 → 0.6740.333 → 0.3240.648 → 0.634
4Seasons0.568 → 0.5660.581 → 0.5590.805 → 0.803

In 10 of these 12 cells the cached loops make the variant look better than its own run does, so the submitted ablation, if anything, understates the contribution of the long pass. For full CLASP, we re-ran the whole pipeline on all 51 sequences with all three backbones (153 runs), verifying loops after the scale solve. Submitted uses the cached loop set and fresh the re-verified loops:

BackboneLong (44): submitted → fresh7-Scenes: submitted → fresh
π³0.789 / 30.77 → 0.791 / 30.600.787 / 0.06 → 0.787 / 0.06
VGGT0.685 / 36.25 → 0.685 / 36.220.693 / 0.11 → 0.712 / 0.11
VGGT-Ω0.811 / 16.53 → 0.812 / 16.530.863 / 0.03 → 0.861 / 0.03

The dataset means barely move and no admitted span changes in any run, so the gains come from the two long-pass roles rather than from the cached loops.

AuthorsOpenReview ↗

Response to Reviewer X5sz — Part 2

Q3, W2: VGGT-Long with the depth-scale reference.

We added the long-pass depth rows of Eq. (2) to the released VGGT-Long as residuals on each chunk's log-scale, keeping its chunks, alignment, loop closure and optimizer. Cells are AUC@30 ↑ / ATE (m) ↓ with π³:

Variant7-Scenes4SeasonsKITTI-360KITTI Odo.TartanAirV. KITTI
VGGT-Long0.680 / 0.0910.595 / 174.10.617 / 51.40.841 / 26.50.954 / 0.1600.827 / 5.2
+ our depth rows0.689 / 0.0880.591 / 130.30.637 / 34.60.846 / 20.30.958 / 0.1510.827 / 1.8
CLASP0.787 / 0.0580.636 / 93.50.721 / 17.10.860 / 18.20.978 / 0.0830.874 / 2.5

The depth rows consistently cut VGGT-Long's ATE across all datasets.The measurement thus transfers to another system and fixes scale there. Against plain VGGT-Long, AUC@30 changes by at most 0.02 because VGGT-Long's chunk alignment fixes the relative poses within each chunk. And CLASP's AUC@30 stays above the augmented VGGT-Long's in all six cells we ran. The scale part of the gain is therefore not tied to our graph.

Q4, W1: Span-level admission, how often it errs, and the rotation prior.

We compared span-level admission with admitting all candidate spans under the same robust weighting, on 7-Scenes, KITTI Odometry and 4Seasons with all three backbones and freshly verified loops. Cells are AUC@30 for admitting all / span-level admission, with the number of spans CLASP admits in brackets:

DatasetCandidate spansπ³VGGTVGGT-Ω
7-Scenes420.789 / 0.787 (25)0.709 / 0.712 (19)0.863 / 0.861 (29)
KITTI Odo.650.583 / 0.867 (7)0.490 / 0.684 (3)0.583 / 0.820 (11)
4Seasons580.448 / 0.636 (0)0.397 / 0.677 (0)0.579 / 0.807 (0)

Span-level admission clearly improves over admitting all on KITTI Odometry and 4Seasons, where most long-span poses of the full pass are wrong and admitting all collapses. And it matches admitting all on 7-Scenes, where it admits about half of the candidates or more. Checked against ground truth, about four in five of the spans it rejects on KITTI Odometry and 4Seasons are indeed less accurate in rotation than the chunk trajectory, with every backbone. On 7-Scenes it does reject usable spans, as the reviewer suspects, but at a cost of at most 0.002 AUC@30. In the other direction, a shared error would let an inaccurate span in, but the admitted spans are accurate in absolute terms: the median admitted span class lies within 0.8° of ground truth in rotation and 2.5° in translation direction, with every backbone.

The persistent rotation prior is a robustly weighted measurement rather than only a gauge. Added to the chunk graph without long-pass input, it raises AUC@30 and lowers RRA on all three datasets with π³, most on 4Seasons (+0.052), where no span is admitted. But it matters little on 7-Scenes once spans are admitted. The revision will describe it as such in Sec. 3.3 and Eq. (7).

AuthorsOpenReview ↗

Response to Reviewer X5sz — Part 3

Q5: Ground-truth chunk scales in Tables 10 and 12.

The ground-truth chunk scales use ground-truth camera poses only so they can be defined on the driving datasets without dense depth. A chunk's scale is the median, over its consecutive frame pairs, of the ground-truth step length divided by the predicted step length in the chunk's initially aligned units. Steps shorter than a quarter of the sequence's median ground-truth step are dropped, and a chunk needs more than five valid steps. A sequence has one free global scale, so the ground-truth and the estimated log-scales are each centred on their median over chunks. The scale error is the median absolute deviation of their difference over chunks, without the 1.4826 factor. We will add this definition to App. A.4.

Q6: Evaluation manifest and configuration record.

Upon acceptance, we will release the code of CLASP, the scripts that run the selected baselines and the full testbed configuration (sequence keys, frame lists and every configuration file), so that each table can be reproduced. The testbed, with CLASP's mean time per sequence and the peak allocated GPU memory of its long pass with VGGT-Ω on one H100 is as:

DatasetSequencesFrames per sequenceVGGT-Ω: s / GiB
KITTI Odometry11271–15911027 / 50.1
KITTI-36012630–1000717 / 41.0
4Seasons10232–1000605 / 38.2
TartanAir6301–612224 / 22.4
Virtual KITTI5233–837253 / 23.1
7-Scenes7500–1000703 / 40.5

Every VGGT-Ω run uses the full long pass so the largest peak is 65 GiB on the longest sequence (1591 frames). Every baseline runs from its public code on the same input frames with the settings of App. A.5. One script scores all methods on the same frames. The admission thresholds τ and τ_t were selected with π³ alone by coarse sweeps (for τ: 1°, 2° and 4°) on 7-Scenes and 4Seasons. There was no held-out split. They were then fixed for every backbone and dataset.

W4: VGGT-SLAM 2.0.

We agree that keyframe-level variables and intra-submap pose constraints are not unique to CLASP, and we do not claim them. In VGGT-SLAM 2.0, a submap's scale follows from its alignment to overlapping or loop-closing submaps. CLASP adds a different measurement source: same-frame depth ratios against a second inference context, which need no overlap or image matching, and long-span pose edges from that context admitted by agreement. We also evaluated VGGT-SLAM 2.0 on all Long (44) sequences with the same VGGT weights and its keyframe selection off so that it poses every frame (AUC@30 ↑ / ATE (m) ↓):

V. KITTITartanAirKITTI Odo.KITTI-3604Seasons
VGGT-SLAM 2.00.869 / 2.440.954 / 0.130.534 / 96.50.267 / 143.40.432 / 267.9
CLASP (VGGT)0.904 / 1.260.969 / 0.110.682 / 35.10.462 / 56.10.674 / 52.9

CLASP is ahead on all five datasets in both metrics. The revision will add this technical comparison to the related work and the table to the appendix.

Reviewer_U5GgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes CLASP, a training-free offline reconstruction framework that combines overlapping chunk predictions and a full-sequence pass from the same frozen backbone through scale recovery and a robust per-frame pose graph. Experiments on 51 sequences from 6 datasets using 3 backbones show substantial ATE reductions over chunk-based baselines on long driving trajectories, while generally matching or improving on full pass accuracy on shorter sequences.

Critical strengths

S1. The paper goes beyond the previously stated accumulated drift in chunk-based reconstruction by showing that full sequence depth can remain a useful scale reference even when its camera predictions deteriorate. The diagnostics in Appendix A.7 provide convincing support for this distinction.

S2. The design follows an intuitive principle: obtain scale from depth and admit long-span poses based on agreement. The same frame depth-ratio constraints require neither camera poses nor image matching, and the method reuses local geometry with per-chunk scale corrections and per-frame pose updates.

S3. The evaluation covers diverse indoor, outdoor, driving, and synthetic environments across three frozen backbones. The strong trajectory results across these settings support the broader applicability of the approach.

Critical weaknesses

[Major] W1. The training-free baselines use their released configurations rather than matched chunk sizes, overlaps, and loop closure settings, leaving uncertainty about how much of the comparative advantage comes from these differences. Although the internal ablations support the proposed components, controlled baseline comparisons would strengthen the attribution of the gains to CLASP’s design.

[Moderate] W2. CLASP retains the computational and memory requirements of a full sequence pass, limiting its scalability despite using chunks for local reconstruction. Appendix A.11 describes a chunk-and-loop fallback, but its performance is not evaluated, leaving the method’s effectiveness under tighter memory constraints unclear.

[Minor] W3. The methodological novelty appears concentrated in separating the full pass’s depth-scale reference from its selectively admitted pose constraints. Accumulated alignment drift and pose graph optimization are established ideas, so I view the contribution primarily as a useful integration of complementary predictions.

Questions for authors

Q1. How do the strongest training-free baselines compare when chunk size, overlap, and loop closure settings are matched to CLASP where applicable?

Q2. How much accuracy is lost when using the chunk-and-loop fallback described in Appendix A.11, and how does it compare with existing chunk-based methods when the full pass cannot run?

Q3. Which datasets or sequences produce the ATE and depth results in Fig. 1, and how are these values aggregated?

4 comments
AuthorsOpenReview ↗

Response to Reviewer U5Gg — Part 1

We thank the reviewer for reviewing our paper and providing valuable comments. We address the concerns below and are open to further discussion and questions. Cells are AUC@30 ↑ / ATE (m) ↓.

W1, Q1: Baselines at matched chunk size, overlap and loop closure.

We have already run VGGT-Long, the baseline closest to CLASP's design, at CLASP's layout since 60/30 (60-frame chunks, 30 frames shared between consecutive chunks) is its default. It is also the only baseline that shares CLASP's chunk structure and produces best results among other training-free methods. We therefore add three controls with the released VGGT-Long to further strengthen the attribution of the gains to CLASP’s design.

  1. Loop closure removed from CLASP only. Cells are VGGT-Long with its loop closure vs. CLASP without loops:
Datasetπ³VGGTVGGT-Ω
7-Scenes0.680 / 0.09 vs. 0.786 / 0.060.667 / 0.10 vs. 0.700 / 0.100.764 / 0.06 vs. 0.862 / 0.03
V. KITTI0.827 / 5.18 vs. 0.874 / 2.500.887 / 1.37 vs. 0.904 / 1.260.889 / 0.71 vs. 0.906 / 0.69
TartanAir0.954 / 0.16 vs. 0.978 / 0.080.950 / 0.16 vs. 0.969 / 0.110.949 / 0.16 vs. 0.975 / 0.12
KITTI Odo.0.841 / 26.5 vs. 0.833 / 21.70.703 / 37.2 vs. 0.653 / 38.60.832 / 23.5 vs. 0.778 / 28.2
KITTI-3600.617 / 51.4 vs. 0.700 / 20.70.351 / 118.3 vs. 0.442 / 58.00.614 / 38.2 vs. 0.660 / 24.8
4Seasons0.595 / 174.1 vs. 0.630 / 93.80.541 / 93.7 vs. 0.650 / 54.40.795 / 31.6 vs. 0.782 / 27.6

Even without loops, CLASP is ahead in AUC@30 in most cells, so its advantage is not a difference in loop handling for most practical situations.

  1. Larger chunks for VGGT-Long. We use π³ and measure AUC@30 at each chunk / overlap:
Dataset60/30120/60240/120CLASP
7-Scenes0.6800.7340.7110.787
V. KITTI0.8270.8390.8630.874
TartanAir0.9540.9530.9350.978
KITTI Odo.0.8410.6990.6170.860
KITTI-3600.6170.4280.1970.721
4Seasons0.5950.5570.4180.636

Larger chunks help VGGT-Long on 7-Scenes and Virtual KITTI, but lower its AUC@30 on TartanAir and sharply on KITTI Odometry, KITTI-360 and 4Seasons. And its best setting stays below CLASP on all six.

  1. Our depth measurements inside VGGT-Long. With π³ as base model, we report the long-pass depth rows of Eq. (2) added to VGGT-Long's optimizer, with its chunks, alignment and loop closure unchanged:
Variant7-Scenes4SeasonsKITTI-360KITTI Odo.TartanAirV. KITTI
VGGT-Long0.680 / 0.0910.595 / 174.10.617 / 51.40.841 / 26.50.954 / 0.1600.827 / 5.2
+ our depth rows0.689 / 0.0880.591 / 130.30.637 / 34.60.846 / 20.30.958 / 0.1510.827 / 1.8
CLASP0.787 / 0.0580.636 / 93.50.721 / 17.10.860 / 18.20.978 / 0.0830.874 / 2.5

The depth rows lower VGGT-Long's ATE on every dataset with superior or comparable AUC@30 numbers, so the measurement fixes scale in another system too. This same-frame depth-ratio measurement is one of our two key technical contributions: on KITTI Odometry, KITTI-360 and 4Seasons it accounts for about half or more of CLASP's ATE gain over VGGT-Long. So CLASP stays ahead in ATE and AUC@30 over VGGT-Long in most cases.

AuthorsOpenReview ↗

Response to Reviewer U5Gg — Part 2

W2, Q2: Accuracy of the chunk-and-loop fallback of App. A.11.

We ran it end to end with all three backbones on all 51 sequences, and compare it with full CLASP run end to end in the same way. AUC@30 for fallback / full CLASP / VGGT-Long:

Datasetπ³VGGTVGGT-Ω
7-Scenes0.719 / 0.787 / 0.6800.739 / 0.712 / 0.6670.784 / 0.861 / 0.764
V. KITTI0.861 / 0.874 / 0.8270.906 / 0.897 / 0.8870.884 / 0.905 / 0.889
TartanAir0.971 / 0.979 / 0.9540.970 / 0.969 / 0.9500.970 / 0.975 / 0.949
KITTI Odo.0.816 / 0.867 / 0.8410.599 / 0.684 / 0.7030.780 / 0.820 / 0.832
KITTI-3600.674 / 0.722 / 0.6170.324 / 0.461 / 0.3510.634 / 0.687 / 0.614
4Seasons0.566 / 0.636 / 0.5950.559 / 0.677 / 0.5410.803 / 0.807 / 0.795

Without the long pass, CLASP loses accuracy against full CLASP on every dataset with π³ and VGGT-Ω. Against VGGT-Long the fallback is ahead in 12 of the 18 cells but roughly stays at the same level. When memory is tight, a long pass on every 4th frame is the better option: on a 4630-frame KITTI-360 drive with π³, it needs under a third of the full pass's memory and is at least as accurate as the full pass, while the fallback loses most of the accuracy.

W3: Novelty concentrated in separating the depth-scale reference from the pose constraints.

We clarify that CLASP contributes by treating a chunk-wise pass and a full-sequence pass of the same frozen model as two measurement sources in one per-frame pose graph, and takes from each only what it measures reliably. Its two technical components are (i) a same-frame depth-ratio measurement that gives every chunk a scale against one global reference, without overlap chaining or image matching, and (ii) a span-level admission test that decides, per sequence, which long-span poses of the full pass to trust. With one fixed setting, this works across three backbones and on both indoor and driving data without training. We see this separation as one part of a broader design which can trigger new designs of future geometric foundation models, rather than the whole contribution. We will state these contributions more clearly in the introduction.

Q3: Datasets and aggregation behind Fig. 1.

Both panels use π³. The ATE panel is the unweighted mean over the 44 sequences of KITTI Odometry, KITTI-360, 4Seasons, TartanAir and Virtual KITTI: 176.78 m for the long pass, 60.82 m for VGGT-Long and 30.77 m for CLASP. The depth panel covers the 11 of these sequences with dense ground-truth depth (TartanAir, Virtual KITTI): AbsRel 0.0896, 0.1091 and 0.0893. The caption will state both.

Thanks to the authors for their detailed rebuttal, which has addressed my concerns. I encourage the authors to include the additional results in the supplementary material of the final version. I will maintain my score of weak accept.

AuthorsOpenReview ↗

Thank you for your response

Dear Reviewer U5Gg,

Thank you very much for your constructive guidance! We will incorporate the new experimental results into the revision as suggested.

Reviewer_n1u8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CLASP, a training-free framework for improving camera trajectories and geometric consistency in long-sequence feed-forward 3D reconstruction. Processing an entire sequence jointly can degrade camera-pose predictions, while assembling independently reconstructed chunks can accumulate pose and scale errors. CLASP combines overlapping chunk passes with a full-sequence pass of the same frozen backbone. The chunks provide local relative motion and output geometry, while same-frame depth ratios between the two passes provide a common scale reference. Long-pass relative-pose constraints are additionally incorporated when they agree sufficiently with a chunk-derived reference trajectory. A robust pose-graph optimization recovers the final camera trajectory.

Critical strengths
  • The experiments demonstrate solid improvements over chunk-based alignment methods, particularly in camera trajectory accuracy on long driving sequences.

  • The method requires no additional training and is evaluated with three reconstruction backbones, supporting its applicability across models.

  • The motivation is clear and supported by diagnostic experiments.It provides an interesting insight for future long-context reconstruction methods.

Critical weaknesses
  • The complete method requires a full-sequence pass and therefore is limited by the backbone’s long-context computational and memory requirements. Appendices A.10–A.11 acknowledge this limitation, but the experiments do not establish performance beyond the reported sequence lengths or evaluate the fallback without long-pass measurements. An evaluation using a sparse set of reference frames could help assess whether the central idea extends to substantially longer sequences.

  • The method section is difficult to follow in places, particularly the distinction between candidate long-pass measurements and the chunk-derived reference trajectory.

Questions for authors
  1. Can the long-pass scale-reference mechanism be extended to substantially longer sequences using selected reference frames or hierarchical processing? How much accuracy is lost when the full long pass is unavailable and the method falls back to chunk and loop constraints?

  2. Appendix A.9 provides useful threshold sensitivity experiments on two datasets with π³. Does the optimal choice change with different model? And for model with different performance capabilities?

  3. As a possible extension, could the complementary chunk and long-pass measurements be combined through a learned fusion module? Which aspects of the current geometric optimization would be most difficult to replace while retaining robustness?

3 comments
AuthorsOpenReview ↗

Response to Reviewer n1u8

We thank the reviewer for suggesting sparse reference frames, which we implemented and report below. Cells are AUC@30 ↑ / ATE (m) ↓, and Long (44) is the mean over the 44 sequences of KITTI Odometry, KITTI-360, 4Seasons, TartanAir and Virtual KITTI.

W1, Q1: Longer sequences, and the accuracy lost without the long pass.

Thank you for your suggestions. We compare the full CLASP model and variants with different strides as below:

Long passπ³VGGTVGGT-ΩTime of the pass
Full0.789 / 30.80.685 / 36.30.811 / 16.5100%
Every 2nd frame0.790 / 31.50.685 / 41.90.817 / 16.727–32%
Every 4th frame0.785 / 35.80.678 / 46.90.814 / 20.38–12%
Every 8th frame0.792 / 40.10.675 / 60.60.817 / 29.02–5%
w/o long pass0.751 / 77.90.611 / 84.20.794 / 24.00%

Against the full long pass, AUC@30 holds at every stride, and ATE degrades mainly at k = 8 but still performs mostly better than the "w/o long pass" variant.

As for longer sequences, on a 4630-frame KITTI-360 drive with π³, 2.9 times the longest sequence in the paper, the full long pass reads 0.653 / 40.8 m at a peak of 72.5 GiB allocated GPU memory. The pass on every 4th frame reads 0.670 / 34.6 m at 22.2 GiB, on every 8th frame 0.680 / 33.5 m at 13.8 GiB, and no long pass 0.326 / 171.1 m. VGGT-Ω's full long pass no longer fits an 80 GiB card at 2000 frames, yet on a 4000-frame drive its pass on every 4th frame reads 0.710 / 19.8 m, against 0.482 / 123.2 m without a long pass.

Removing the long pass (last row above) lowers Long (44) from 0.789 / 30.8 to 0.751 / 77.9 m with π³. Run end to end, the chunk-and-loop fallback of App. A.11 loses AUC@30 against full CLASP in 15 of the 18 dataset–backbone cells, by up to 0.14 (KITTI-360 with VGGT), which puts it roughly at VGGT-Long's level. Our reply to Reviewer U5Gg (Part 2) gives the table.

W2: The method section is difficult to follow.

Thank you for pointing this out. We will make the discussions about the reference trajectory and candidate measurements more clear in the revision.

Q2: Does the optimal threshold change with the model?

We repeated the sweeps of App. A.9 with all three backbones. Within a factor of 2.5 of the adopted τ = 1°, the worst loss against τ = 1° on any dataset is 0.017 AUC@30 with π³, 0.013 with VGGT and 0.004 with VGGT-Ω; for τ_t it is at most 0.008. On KITTI Odometry, KITTI-360 and 4Seasons the thresholds hardly matter (at most 0.003), since few spans there are close to admission. The optimum does shift for the weakest backbone, VGGT: a looser τ = 2° would gain 0.012 on 7-Scenes but lose 0.011 on TartanAir. The test also adapts by itself, admitting 49, 73 and 80 of 294 candidate spans with VGGT, π³ and VGGT-Ω. The thresholds were chosen with π³ and then fixed, which the revision will state.

Q3: A learned fusion module.

The current scale reference is a closed-form measurement with nothing to learn. Admission is where learning could help. Checked against ground truth, most spans rejected on 7-Scenes were usable (with π³, 16 of 17 were at least as accurate in rotation as the chunk-derived reference trajectory), so a learned confidence could recover them and could also guard the always-on rotation prior. The hardest properties to keep would be generalisation over sequence length and backbone, since one setting now covers three backbones and 200 to 4630 frames, and global consistency under loop closure. We will add this discussion to the revision.

Thank you for the additional explanation and experiments. My concerns have been fully addressed. I will maintain my score of weak acceptance.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer n1u8,

Thank you very much for your constructive guidance. We promise that we will follow your suggestions and put all new results into the final revision after proofreading.

Reviewer_N5ixOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

CLASP combines overlapping chunks and a full-sequence pass of the same frozen reconstruction model. Chunks provide local motion and geometry; same-frame depth ratios provide a shared scale reference, while long-span poses are admitted when they agree with a chunk-derived trajectory. A robust pose graph fuses these measurements. Experiments cover three backbones and six datasets, with the clearest gains on driving trajectories.

Critical strengths
  1. The use of long-pass depth as an overlap-independent scale reference is convincing. In the π³ ablation in Table 12, removing it raises 4Seasons ATE from 93.51 to 242.15 m. Table 10 also directly checks scale accuracy when long-pass poses are poor.
  2. The evaluation is reasonably broad and includes meaningful counterexamples. With VGGT-Ω, KITTI-360 ATE drops from 38.18 m for VGGT-Long to 20.26 m (Table 1). The appendix reports all three backbones and shows cases where CLASP remains weaker, rather than implying uniform improvements.
Critical weaknesses
  1. The evaluated pipeline does not fully follow the stated loop-verification order. Section 3 and Appendix A.2 verify loops using scale-corrected points, whereas Appendix A.5 reuses loops from an earlier run and keeps them fixed during the new scale solve. The reported agreement with a fresh run is reassuring, but covers only π³ on 4Seasons. Since Figure 7 identifies loop errors as a failure source, this discrepancy matters for the main comparisons.
  2. The accuracy improvement carries a substantial resource cost: 122.0 s and 24.82 GiB versus 34.8 s and 7.49 GiB for VGGT-Long on the 964-frame π³ comparison in Appendix A.10. The full pass retains quadratic attention, and longer recordings are split for evaluation. The paper acknowledges this limitation, but does not establish whether the gains remain attractive against a chunk-based baseline given the same inference budget.

References and writing (RW)

Some ablation descriptions need clarification. Section 3.3 says admitted long-pass edges constrain rotations and translations, but Table 4 and Appendix A.5 describe restoring only translations, while rotation error also changes. Please specify which edges and priors remain in each variant and which quantities are recomputed. Figure 1 should also identify the evaluation subset and averaging behind its headline numbers; these are currently difficult to reconcile with the per-dataset tables.

Questions for authors
  1. Can you provide fresh end-to-end results, with loops reverified after scale correction, for the remaining backbone/dataset combinations? Which settings change most relative to the cached-loop results?
  2. At approximately CLASP's inference budget, how much of the advantage remains over VGGT-Long with a larger chunk or overlap? A small controlled comparison would clarify the value of the second context.
3 comments
AuthorsOpenReview ↗

Response to Reviewer N5ix

We thank the reviewer for checking the appendix against the method section. We address the concerns below and are open to further discussion and questions. Both the loop-order point and the ablation-wording point are correct. Cells are AUC@30 ↑ / ATE (m) ↓, and Long (44) is the mean over the 44 sequences of KITTI Odometry, KITTI-360, 4Seasons, TartanAir and Virtual KITTI.

W1, Q1: Fresh end-to-end results with loops re-verified after scale correction.

We ran the pipeline in the order of Sec. 3 on all 51 sequences with all three backbones. Submitted uses the cached loop set of the paper, and fresh re-verifies the loops after the scale solve:

BackboneLong (44): submitted → fresh7-Scenes: submitted → fresh
π³0.789 / 30.77 → 0.791 / 30.600.787 / 0.06 → 0.787 / 0.06
VGGT0.685 / 36.25 → 0.685 / 36.220.693 / 0.11 → 0.712 / 0.11
VGGT-Ω0.811 / 16.53 → 0.812 / 16.530.863 / 0.03 → 0.861 / 0.03

Both loop sets give nearly the same accuracy, and no admitted span changes in any run. The settings that change most are 7-Scenes with VGGT (AUC@30 0.693 → 0.712) and KITTI Odometry with π³ (0.860 → 0.867). The latter is almost entirely sequence 07, the loop failure of Fig. 7: verified after the scale solve, it keeps 294 loops instead of 60, and its ATE falls from 11.2 m to 3.5 m.

W2, Q2: The advantage at CLASP's inference budget.

Thank you for your suggestions. To further validate the empirical advantages of CLASP, we evaluate VGGT-Long with larger chunks. Here "60/30" denotes chunks of 60 frames in which consecutive chunks share 30 frames; it is VGGT-Long's default and also CLASP's chunk layout. We doubled and quadrupled both numbers. We use the released VGGT-Long with loop closure, π³ and AUC@30 and have results as:

Dataset60/30120/60240/120CLASP
7-Scenes0.6800.7340.7110.787
V. KITTI0.8270.8390.8630.874
TartanAir0.9540.9530.9350.978
KITTI Odo.0.8410.6990.6170.860
KITTI-3600.6170.4280.1970.721
4Seasons0.5950.5570.4180.636

Larger chunks help VGGT-Long on 7-Scenes and Virtual KITTI, but lower its AUC@30 on TartanAir and sharply on KITTI Odometry, KITTI-360 and 4Seasons. On these three, at 240/120 it uses about CLASP's memory (about 19.5 GiB against 19 to 28 GiB). Its best setting stays below CLASP on all six datasets. At 120/60 with VGGT and VGGT-Ω on these three datasets, it is slightly ahead of CLASP only with VGGT-Ω on KITTI Odometry (0.827 / 19.0 m against 0.820 / 21.0 m).

Additionally, with the long pass on every 8th frame, the pass takes 2% of the full pass's time and its peak allocated GPU memory is 7.8 GiB on average (10.4 GiB at most), comparable to the 8.9 GiB of the released VGGT-Long. CLASP then keeps the AUC@30 of the full long pass at a somewhat higher ATE but stays well ahead of VGGT-Long on Long (44) (0.792 / 40.1 m against 0.738 / 60.8 m).

RW: Edges and priors in each ablation variant.

We clarify that admitted edges constrain both rotations and translations, and Table 4 and App. A.5 incorrectly describe the last step as restoring translations only. Our response to Reviewer 1Mzy (Part 1, W1, Q1) gives the corrected Table 4, which adds the long-pass inputs one at a time and states which edges and priors each row contains. In every row, the scale solve, the rescaling of link and loop translations to the new chunk scales, and the rotation and position solves are recomputed. The predictions of both passes and the loop set are held fixed.

RW: Evaluation protocol in Fig. 1.

Both panels use π³ as base model. The ATE panel is the unweighted mean over the 44 sequences of Long (44): 176.78 m for the long pass, 60.82 m for VGGT-Long and 30.77 m for CLASP. The depth panel covers the 11 of these sequences with dense ground-truth depth (TartanAir, Virtual KITTI): AbsRel 0.0896, 0.1091 and 0.0893. The caption will state both.

Reviewer_N5ixOpenReview ↗

Thank you for your response

Thanks for the response. The new results address my main concerns about loop verification and compute cost. Please include these results and clarifications in the revision. I’ll keep my weak accept score.

Dear Reviewer N5ix,

Thank you very much for your constructive guidance! We will incorporate the results and your suggestions into the revision.

Reviewer_1MzyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

CLASP combines chunk-wise and full-sequence predictions from a frozen 3D reconstruction model, using global depth to correct chunk scales and selectively admitting long-range pose constraints to improve reconstruction across three backbones and six datasets.

Critical strengths
  1. The method makes good use of the different strengths of short-chunk and full-sequence predictions without requiring additional training.

  2. With the same backbone, CLASP achieves lower ATE than VGGT-Long in all nine driving dataset–backbone combinations.

  3. The scale ablation clearly supports the main idea: removing full-sequence depth measurements increases ATE from 93.51 m to 242.15 m on 4Seasons with π³.

Critical weaknesses
  1. It is unclear how much each type of global pose information contributes, since the ablations do not clearly separate long-range rotations, translations, and the rotation prior.

  2. Most results reuse cached loop closures, and the paper checks their agreement with a fresh full run only for π³ on 4Seasons.

  3. The threshold tests show stable results, but do not show whether the proposed filtering works better than simply keeping or dropping all long-range pose edges.

  4. Runtime and memory are measured on only one sequence, so it is hard to judge how the extra cost grows with sequence length.

Questions for authors
  1. Which rotation constraints remain in each ablation, and how does performance change when all full-sequence pose information is removed?

  2. On a few representative sequences, does recomputing loop closures change the accepted loops or final accuracy compared with using cached loops?

  3. How does the proposed filtering compare with keeping all or dropping all candidate long-range pose edges?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1Mzy — Part 1

We thank the reviewer for reviewing our paper and providing valuable comments. We address the concerns below and are open to further discussion and questions. Cells are AUC@30 ↑ / ATE (m) ↓ with π³ unless stated. Long (44) is the mean over the 44 sequences of KITTI Odometry, KITTI-360, 4Seasons, TartanAir and Virtual KITTI.

W1, Q1: Which rotation constraints remain, and what happens without long-pass pose information.

The texts in the paper are unclear. Actually, the "Chunk graph" row keeps the long-pass rotation prior (Eq. 7), and the last row in Tab. 4 of the main paper adds the admitted spans to both the rotation and the position solve, not only translations. The following table adds the long-pass inputs one at a time, with the predictions and loop set of Tab. 4. Cells are AUC@30 ↑ / ATE (m) ↓ / RRA (°) ↓ / RTA (°) ↓:

Variant7-Scenes4SeasonsKITTI-360KITTI Odo.TartanAirV. KITTI
Chunk graph, no long-pass input0.718 / 0.084 / 4.37 / 9.860.568 / 245.4 / 11.54 / 13.960.679 / 39.3 / 9.29 / 8.190.824 / 43.3 / 3.99 / 4.980.972 / 0.113 / 0.70 / 1.170.866 / 4.7 / 0.90 / 9.64
+ rotation prior (Table 4 "Chunk graph")0.735 / 0.081 / 3.72 / 9.410.620 / 242.1 / 9.43 / 12.260.698 / 37.8 / 7.61 / 8.070.846 / 43.4 / 3.33 / 4.680.974 / 0.110 / 0.62 / 1.130.872 / 4.6 / 0.68 / 9.60
+ long-pass scale0.734 / 0.081 / 3.72 / 9.370.636 / 93.5 / 9.43 / 10.810.721 / 17.1 / 7.61 / 6.690.863 / 17.1 / 3.33 / 3.820.976 / 0.096 / 0.62 / 1.070.876 / 1.9 / 0.68 / 9.49
+ admitted spans (full CLASP)0.787 / 0.058 / 3.61 / 7.420.636 / 93.5 / 9.43 / 10.810.721 / 17.1 / 7.61 / 6.690.860 / 18.2 / 3.36 / 4.040.978 / 0.083 / 0.60 / 0.970.874 / 2.5 / 0.71 / 9.43
CLASP without long-pass pose input (scale kept)0.716 / 0.083 / 4.37 / 9.830.587 / 98.8 / 11.54 / 11.840.695 / 24.9 / 9.29 / 7.690.845 / 17.4 / 3.99 / 3.940.974 / 0.100 / 0.70 / 1.110.869 / 2.0 / 0.90 / 9.53

Every row contains the chunk links and verified loops. The rotation prior enters in the second row, and the rotation part of the admitted spans only in CLASP. The last row answers Q1: it removes the prior and both parts of the admitted spans from CLASP and keeps only the long-pass scale.

Compared with full CLASP, removing all long-pass pose input (last row) lowers AUC@30 on every dataset, most on 7-Scenes (−0.07) and 4Seasons (−0.05). On KITTI Odometry, KITTI-360, 4Seasons and Virtual KITTI most of CLASP's ATE gain over the first row remains, because the long-pass scale carries it. Against the first row, the rotation prior (second row) raises AUC@30 and lowers RRA on all six datasets. On 7-Scenes most of the gain comes instead from the admitted spans: without them (third row), ATE stays at the level of the first row. About 60% of their AUC@30 gain comes from the loop translation noise that an admitted span sets (App. A.2) and most of the rest from their translation part, while their rotation part alone accounts for the RRA change. All three long-pass inputs therefore contribute: the rotation prior raises AUC@30 on all six datasets, the long-pass scale cuts ATE on every dataset except indoor 7-Scenes and the admitted spans give most of the 7-Scenes gain (AUC@30 0.734 → 0.787). Together they improve on the first row in both AUC@30 and ATE on all six datasets. The revision will correct Tab. 4, App. A.5 and the description of the prior in Sec. 3.3 and Eq. (7) accordingly.

W2, Q2: Do fresh loop closures change the accepted loops or the accuracy?

We re-ran the full pipeline on all 51 sequences with all three backbones (153 runs), verifying loops after the scale solve. Submitted uses the cached loop set of the paper and fresh the re-verified loops:

BackboneLong (44): submitted → fresh7-Scenes: submitted → fresh
π³0.789 / 30.77 → 0.791 / 30.600.787 / 0.06 → 0.787 / 0.06
VGGT0.685 / 36.25 → 0.685 / 36.220.693 / 0.11 → 0.712 / 0.11
VGGT-Ω0.811 / 16.53 → 0.812 / 16.530.863 / 0.03 → 0.861 / 0.03

In these runs, the loop sets do change: verified after the scale solve, more loops pass. But the accuracy barely does: no admitted span changes in any run, and the mean AUC@30 moves by at most 0.002 on Long (44) and 0.02 on 7-Scenes for every backbone. The conclusions of the paper therefore hold with re-verified loops.

AuthorsOpenReview ↗

Response to Reviewer 1Mzy — Part 2

W3, Q3: Filtering against keeping or dropping all long-span edges.

Each row is the mean over that dataset's sequences in the paper, with freshly verified loops and the same robust weighting in every column. Cells are AUC@30 for drop all / span admission (CLASP) / keep all:

Datasetπ³VGGTVGGT-Ω
7-Scenes0.758 / 0.787 / 0.7890.677 / 0.712 / 0.7090.811 / 0.861 / 0.863
V. KITTI0.876 / 0.874 / 0.8730.884 / 0.897 / 0.8870.885 / 0.905 / 0.902
TartanAir0.976 / 0.979 / 0.9760.970 / 0.969 / 0.9120.973 / 0.975 / 0.970
KITTI-3600.722 / 0.722 / 0.4740.461 / 0.461 / 0.4300.687 / 0.687 / 0.416
KITTI Odo.0.870 / 0.867 / 0.5830.683 / 0.684 / 0.4900.822 / 0.820 / 0.583
4Seasons0.636 / 0.636 / 0.4480.677 / 0.677 / 0.3970.807 / 0.807 / 0.579

On KITTI Odometry, KITTI-360 and 4Seasons, dropping all spans changes almost nothing: CLASP admits no span on KITTI-360 and 4Seasons, so the two are the same solution, and only a few on KITTI Odometry. Where spans are admitted, dropping them costs accuracy against CLASP: on 7-Scenes AUC@30 falls by 0.03 to 0.05 with every backbone and ATE rises by 20% to 50%, and on Virtual KITTI AUC@30 falls by up to 0.02 with VGGT and VGGT-Ω. Keeping all spans, on the other hand, collapses on KITTI Odometry, KITTI-360 and 4Seasons even with robust weighting (4Seasons with π³: ATE about 3.5 times that of CLASP). Over these long drives the full-sequence pass is itself inaccurate, and its long-span poses are wrong as a rule, not as exceptions. Checked against ground truth, the typical candidate span there is far worse than the chunk trajectory, above all in translation direction. A robust kernel down-weights isolated outliers, but when most edges are wrong they outweigh the chunk links and pull the positions off course. Span admission rejects these spans on the long drives and keeps them indoors without a dataset-specific switch.

W4: Cost against sequence length.

We thank the reviewer for pointing out that the paper measures runtime and memory on a single sequence. We now measure how the cost grows with sequence length, for all three backbones, on one H100, as the long pass's seconds / peak allocated GPU memory (GiB). Columns give the number of input frames. Driving frames are wide and enter π³ and VGGT at 518×140 (KITTI-360) or 518×154 (KITTI Odometry), about 400 tokens per frame. And VGGT and VGGT-Ω keep the layer outputs their heads need in a memory-mapped cache off the GPU. Besides the full pass, each backbone runs the pass on every k-th frame (k = 2, 4, 8), which still supplies every input CLASP takes from the long pass:

Long pass500100020004000
π³, every frame13 / 13.347 / 21.3150 / 34.4608 / 63.4
π³, every 2nd4.0 / 9.413 / 13.338 / 19.9151 / 34.4
π³, every 4th1.3 / 7.44.5 / 9.410 / 12.639 / 19.9
π³, every 8th0.5 / 6.41.7 / 7.43.2 / 9.011 / 12.6
VGGT, every frame39 / 14.0104 / 21.9250 / 35.0919 / 64.0
VGGT, every 2nd15 / 10.038 / 14.085 / 20.5268 / 35.0
VGGT, every 4th7.0 / 8.016 / 10.033 / 13.392 / 20.5
VGGT, every 8th3.3 / 7.08.5 / 8.011 / 9.632 / 13.3
VGGT-Ω, every frame129 / 24.6412 / 43.2out of memoryout of memory
VGGT-Ω, every 2nd52 / 18.2133 / 24.6410 / 43.2out of memory
VGGT-Ω, every 4th22 / 12.152 / 18.2230 / 24.6394 / 43.2
VGGT-Ω, every 8th8.3 / 9.025 / 12.144 / 18.2136 / 24.6

Memory grows about linearly and time about quadratically with the number of frames. A pass on every k-th frame costs about as much as a full pass over 1/k of the frames. At 4000 frames with π³, every 4th frame takes about a third of the full pass's memory and a fifteenth of its time. The stride also lets VGGT-Ω, whose full pass no longer fits at 2000 frames, run at 4000 frames, where it lowers ATE about sixfold against running without a long pass. The accuracy is kept: on Long (44) with π³, AUC@30 stays within 0.005 of the full pass at every stride, while ATE grows with the stride, most at k = 8. Since the chunk passes grow linearly with length (App. A.10), the superlinear part of CLASP's cost is the long pass, and the stride keeps it in check. On every 4th frame, the pass over 4000 frames costs less than the full pass over 2000 frames with π³ and VGGT, and it is what lets VGGT-Ω reach 4000 frames at all.

I appreciate the authors for providing extensive supplementary experiments regarding robust graphs, span admission, and efficiency. In particular, the experiments on span admission confirm the method's strong compatibility across different scenarios.

I decide to raise my score towards acceptance.

AuthorsOpenReview ↗

Thank you for your response

Dear Reviewer 1Mzy,

Thank you again for your helpful feedbacks and raising your score! We are very glad our additional experimental results addressed your questions. We will incorporate these results and your other suggestions into our final paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.