- 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
-
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.
-
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?
-
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.
-
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.
-
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.
-
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
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:
| Variant | 7-Scenes | 4Seasons | KITTI-360 | KITTI Odo. | TartanAir | V. KITTI |
|---|---|---|---|---|---|---|
| VGGT-Long | 0.680 / 0.091 | 0.595 / 174.1 | 0.617 / 51.4 | 0.841 / 26.5 | 0.954 / 0.160 | 0.827 / 5.2 |
| + our depth rows | 0.689 / 0.088 | 0.591 / 130.3 | 0.637 / 34.6 | 0.846 / 20.3 | 0.958 / 0.151 | 0.827 / 1.8 |
| CLASP | 0.787 / 0.058 | 0.636 / 93.5 | 0.721 / 17.1 | 0.860 / 18.2 | 0.978 / 0.083 | 0.874 / 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:
| Dataset | Candidate spans | π³ | VGGT | VGGT-Ω |
|---|---|---|---|---|
| 7-Scenes | 42 | 0.789 / 0.787 (25) | 0.709 / 0.712 (19) | 0.863 / 0.861 (29) |
| KITTI Odo. | 65 | 0.583 / 0.867 (7) | 0.490 / 0.684 (3) | 0.583 / 0.820 (11) |
| 4Seasons | 58 | 0.448 / 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).
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:
| Dataset | Sequences | Frames per sequence | VGGT-Ω: s / GiB |
|---|---|---|---|
| KITTI Odometry | 11 | 271–1591 | 1027 / 50.1 |
| KITTI-360 | 12 | 630–1000 | 717 / 41.0 |
| 4Seasons | 10 | 232–1000 | 605 / 38.2 |
| TartanAir | 6 | 301–612 | 224 / 22.4 |
| Virtual KITTI | 5 | 233–837 | 253 / 23.1 |
| 7-Scenes | 7 | 500–1000 | 703 / 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. KITTI | TartanAir | KITTI Odo. | KITTI-360 | 4Seasons | |
|---|---|---|---|---|---|
| VGGT-SLAM 2.0 | 0.869 / 2.44 | 0.954 / 0.13 | 0.534 / 96.5 | 0.267 / 143.4 | 0.432 / 267.9 |
| CLASP (VGGT) | 0.904 / 1.26 | 0.969 / 0.11 | 0.682 / 35.1 | 0.462 / 56.1 | 0.674 / 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.
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:
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):
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:
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.