acceptodds
Under review as a conference paper at ICLR 2027

Many Eyes, One World: Feed-Forward 3D Reconstruction from Mixed Cameras

Abstract

Real-world capture is heterogeneous: perspective, fisheye, and 360 panoramic images can coexist within a single reconstruction task, yet most feed-forward 3D reconstruction models assume perspective imagery and a uniform input representation. Recent models handling several camera types are either informed of the camera type for each view or reconstruct one image pair at a time. No single-pass method reconstructs mixed-camera tuples containing full panoramas from images alone. We present MEOW, a feed-forward system that jointly reconstructs metric pointmaps and camera poses from one -view tuple mixing perspective, fisheye and full-panorama images, in a single forward pass from images alone: no calibration, distortion parameters, camera-type labels or poses are supplied for any view. Our guiding design philosophy is to treat heterogeneous-camera reconstruction as a data-adaptation problem rather than an architectural redesign. MEOW retains a perspective-pretrained backbone and learns heterogeneous cameras entirely from a procedural data engine, which renders each scene across a continuous manifold of camera models with exact rays and depth, and certifies covisibility for every training tuple. Trained on synthetic tuples only, MEOW transfers zero-shot to real captures: on heterogeneous 2D3DS tuples it achieves 79.9 mAA@30 against 53.8 for Wid3R given the camera type of every view; on our laser-scanned mixed-camera benchmark it registers every four-view mixed tuple with 79.4 AUC@30. The data engine, benchmark, and complete evaluation pipeline will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
feed-forward 3D reconstruction, heterogeneous cameras, camera-agnostic, multi-view pose estimation, metric point maps, synthetic data engine

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_p5zbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors provided an effective pipeline, MEOW, to acheive the feed-forward 3D reconstruction from mixed and unconstrained camera types, without the need of the related camera parameters, such as intrinsics, distortion parameters and so on. Besides, the authors also proposed a laser-scanned benchmark, which is beneficial for a heterogeneous-camera reconstruction community.

The proposed procedural data engine, LENSCOPE, firstly generates diverse scenarios and corresponding panoramas and depth maps, and then sample the mixed-camera views with sufficient covisibility. Meanwhile, MEOW leverages the synthetic camera distrbution fine-tuning on MayAnything to learn the implicit correspondence between different projection models and camera rays, achieving the feed-forward 3D generation with the mixed camera type.

Critical strengths

a) The high-quality synthestic multiple camera type data is providede from the proposed Lenscope and the authors specifically designed the camera sampling, covisibility verification and accurate geometric supervision for sufficient fine-tuning.

b) The images, from different projection model, could be directly used as the input without any porjection transformation. The designs of resizeing without cropping and aspect-ratio embedding subtly injuce the camera projection information into the model.

c) The performance on both the 2D3DS benchmark and the self-captured benchmark demonstrated the superior capability of MEOW.

Critical weaknesses

Although the conducted experiments in this paper show that MEOW is an promising data-centric adaptation framework for camera-agnostic feed-forward 3D reconstruction, I have the following major concerns:

  1. Less convinced technical novelty. After carefully reading this work, I found the main contribution is the proposed data engine and data preparison. While the network architecture is largely inherited from MapAnything, the simple finetuning on the synthestic data already brought the obvious performance improvement, as shown in Tab.8, Appendix and the following training with the addtional two stages even decreased the performance of Acc. and Comp.. Especially, combined with the ablation study in Tab.4, the reason of final performance improvement is unclear, where there are over 800-epoch training on one branch in stage 3. More controlled experiments are required, such as just continuously finetuning the MapAnything with same budgets.

  2. Lack of comparison with the closest baseline, CAM3R. I recognized that CAM3R didn't release the multi-view weights, but the exclusion is somewhat sufficient. CAM3R has already reported multi-view reconstruction results on the 2D3DS dataset, the authors could potentially evaluate MEOW under CAM3R's published evaluation protocol and compare against its reported results.

  3. Insufficient experimental justification for the motivation. Since authors viewed direct heterogeneous-camera inputs as the motivation and highlighted the limitations of rectifying wide-angle images to perspective views (L68-74), a direct comparison with a perspective-conversion pipeline based on MayAnything could help clarify the motivation (the calibration might be estimated by some existing methods, such as AnyCalib[1], CalibAnyView[2]).

Minor Concerns:

  1. More generalization results on real-world data. In this work, some camera types are sampled from panoramas. More experiments with the heterogeneous capture could better explain the robustness. For example, the fisheye images are from some ego-glasses, and the perspective images are from phones or GoPro.
  2. The limitations in multi-views and out-of-scope perspective videos . More discussion for the results in Tab12. and Tab.14 could better clairfy the performance.

[1] Tirado-Garín, Javier, and Javier Civera. "AnyCalib: On-manifold learning for model-agnostic single-view camera calibration." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025. [2] Li, Boying, et al. "Calibanyview: Beyond single-view camera calibration in the wild." arXiv preprint arXiv:2605.14615 (2026).

Questions for authors

See weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer p5zb

Response to Reviewer p5zb

W1. Contribution and the requested control. MEOW is a data-centric adaptation, as the reviewer says and as the paper states: heterogeneous-camera reconstruction is treated as a data problem rather than an architectural one. The contribution is the engine that makes the adaptation possible and its result, the first model to reconstruct mixed tuples with full panoramas in one pass from images alone; extending pretrained models through data is an established route to new capability [1,2]. The requested control is below. Runs A and C start from the same Stage-1 checkpoint with identical scenes, seed and 6,240 updates.

Checkpoint2D3DS mAA@30, full-FoV / croppedLaser mixed AUC@30, full-FoV
MapAnything, public15.4 / 16.419.7
Stage 1, start of the runs42.2 / 68.358.3
A: continued fine-tuning only56.6 / 63.274.5
C: full recipe60.9 / 62.872.5
Released weights80.4 / 77.179.4

With the architecture unchanged, every checkpoint adapted on LENSCOPE is far above the MapAnything it descends from. The added components contribute at every budget: +4.3 at 15 epochs; 76.6† with the aspect-ratio embedding against 70.6† without it after 90 Stage-3 epochs; and 23.2 lost when the embedding is withheld from the released weights. The 859 epochs are history, not a requirement: 76.6† after 90 epochs, 80.6† after 859. On Acc. and Comp., Table 8 was unclear: it compared Stage 1 under its crop loader with later stages under full-field-of-view resizing. Under one loader accuracy improves from 0.169 m to 0.142 m and completeness stays within 0.02 m.

W2. CAM3R. CAM3R released a checkpoint on 20 October, after the submission. It covers the two-view network, without the multi-view alignment or the preprocessing that defines its test sets, so its Table 2 cannot be matched input for input. Under its scoring MEOW reaches 81.2 mAA@30 on 2D3DS tuples of up to eight views and 53.7 over all 88 tuples including those of 15–24 views; CAM3R reports 73.5 on its own image sets, with global alignment and after training on Areas 1–4 of the same dataset. On identical image pairs in both input orders:

mAA@30CAM3RMEOW
Laser benchmark, 288 ordered pairs17.877.5
Real cameras, 130 ordered pairs29.872.7
2D3DS, 138 ordered pairs with short baselines70.088.6
of which panorama–panorama95.384.9

MEOW is ahead on all three sets, by 43 and 60 points on the two that neither model has seen; CAM3R is ahead on panorama pairs of 2D3DS, its training dataset.

W3. Perspective conversion. Lines 68–74 were too brief. We built the pipeline: virtual perspective views covering the full field of view, MapAnything (CC BY-NC v1.1), and ground-truth or AnyCalib calibration with the camera family given. CalibAnyView's multi-view mode assumes intrinsics shared across views, which mixed lenses violate.

Point error without scale fitting: on the laser mixed track, 0.98 m for MEOW against 0.82 m and 0.83 m for conversion with ground-truth and AnyCalib calibration; on 2D3DS, a mean of 0.87 m against 8.46 m and 18.4 m. Given the camera model of every image, which MEOW does not need, conversion gains 0.16 m on the laser office, loses on 2D3DS, and takes five times as long.

Minor 1. Real heterogeneous capture. We added the requested evaluation using independently calibrated physical cameras. In a laser-scanned office we calibrated and localised five physical cameras. We had no smart glasses, so the fisheye source is a calibrated action camera, a GoPro HERO4 Session. Perspective views come from two phones of different makes, an iPhone 12 Pro Max and a Xiaomi 17 Pro with its main and ultrawide lenses, and from a Fujifilm X100VI; panoramas come from an Insta360 X3. On mixed tuples built as in the mixed track of the paper, MEOW reaches 70.5 AUC@30 against 29.9 for the best baseline.

Minor 2. Tables 12 and 14. Both tables mark the range of the adaptation data, room-scale tuples of 2–8 views, as Section 5 states. In Table 14 MEOW leads Wid3R on 79 of the 88 tuples and at every size up to 14 views, beyond that range (67.9 against 56.9 at 9–14 views); only the four tuples of 15–24 views fall behind. Table 12 is video with centimetre baselines: every model solves rotation (RRA@30 of 99–100), and the differences lie in the translation direction of very small motions. At room-scale baselines MEOW stays within 1.1 AUC@30 of its backbone on perspective input (85.7 against 86.8, Table 3).

† Submission ground truth; other 2D3DS values follow an axis-convention fix in the ground truth (every method +0.1 to +0.5 mAA@30, no ranking change, weights unchanged).

[1] Duan et al., Fisheye3R, ECCV 2026. [2] Yang et al., Depth Anything V2, NeurIPS 2024.

Thanks for the detailed response and the additional experiments. They address much of my original concern.

My main remaining question is how much the full training recipe adds beyond continued fine-tuning on LENSCOPE. At the same 15-epoch budget, the full recipe improves full-FoV 2D3DS mAA from 56.6 to 60.9, but scores slightly lower on cropped 2D3DS and drops from 74.5 to 72.5 on the laser mixed track, which also uses full-FoV inputs. This leaves it unclear whether the full recipe offers a consistent benefit over continued fine-tuning on LENSCOPE.

Could you extend both Run A and Run C from the same Stage-1 checkpoint to a longer matched budget, for example 90 epochs, and evaluate them under the same settings? This would help clarify how much the full recipe adds beyond continued fine-tuning alone.

AuthorsOpenReview ↗

**Matched-budget comparison at 60 epochs**

We are glad that the earlier responses addressed much of the reviewer's original concern. The requested comparison is reported below.

Both runs were extended from their 15-epoch checkpoints with the same seed, scenes and hardware. Entries are A / C under the settings of the earlier table, re-evaluated in the same environment.

Epochs2D3DS, full FoV2D3DS, croppedLaser mixed
1556.65 / 60.9363.18 / 62.8974.55 / 72.39
6060.74 / 68.2064.38 / 70.0176.22 / 78.11

At 60 epochs the full recipe is ahead in all three settings, by 7.5, 5.6 and 1.9. Training continues, and we will report results at 90 epochs within the discussion period.

For context on the longer schedule: under the same protocol, the historical checkpoints of the released model change by about four points on full-FoV 2D3DS between 90 and 859 epochs, while laser-panorama AUC@30 rises from 47.9 to 73.0. The matched-budget runs A and C assess the contribution of the full recipe separately.

AuthorsOpenReview ↗

**Update: matched-budget comparison at 90 epochs**

Both runs have now reached 90 epochs under the same conditions as before. Entries are A / C; the earlier rows are repeated for reference.

Epochs2D3DS, full FoV2D3DS, croppedLaser mixed
1556.65 / 60.9363.18 / 62.8974.55 / 72.39
6060.74 / 68.2064.38 / 70.0176.22 / 78.11
9063.19 / 68.7465.69 / 72.0777.76 / 79.66

In our observation, both runs largely converge around 70+ epochs. At 90 epochs the full recipe remains ahead in all three settings, by 5.6, 6.4 and 1.9. The cropped and laser mixed margins are close to those at 60 epochs (5.6 and 1.9). The full-FoV margin is smaller than at 60 epochs (7.5), in line with the checkpoint-to-checkpoint variation noted earlier.

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

MEOW predicts camera poses and metric pointmaps in a single forward pass from tuples that mix perspective, fisheye, and full-panorama images. It requires neither calibration nor camera-type labels. The method fine-tunes a MapAnything backbone on synthetic tuples from LENSCOPE, a procedural indoor engine that supports multiple camera models and rechecks covisibility after camera sampling. It also preserves the full field of view, encodes aspect ratio, uses circular padding for panoramas, and introduces sphere-aware losses. Experiments on 2D3DS and a new laser-scanned office benchmark show substantial gains over prior feed-forward reconstructors on mixed-camera inputs, including camera-aware methods.

Critical strengths

S1. Large and robust gains on mixed-camera tuples. On 2D3DS, MEOW substantially outperforms camera-aware Wid3R.

S2. A clear problem setting and a useful benchmark. The paper isolates a concrete gap: single-pass reconstruction from image-only tuples that include full panoramas. It contributes a laser-scanned benchmark with mixed- and single-camera tracks, scored with the same tuples, ground truth and scorer for every baseline.

Critical weaknesses

Transfer to real heterogeneous cameras is not quantitatively evaluated.** The motivation is real capture that mixes perspective, fisheye and panoramic cameras. This is a reasonable design choice, but it means each benchmark uses one capture device and colour pipeline and an exact, ideal projection, so lens-model error, per-device colour response and rolling shutter are never tested. The only quantitative native-camera data is ADT in Tab. 12. But there's only one camera type per set, the paper does not say whether the pinhole frames are rectified as well.

Questions for authors
  1. MEOW trails Wid3R on all three single-camera laser tracks and also trails its MapAnything backbone on the pinhole and fisheye tracks (Tab. 3). Because Wid3R trains with only one camera class per sample, could part of the mixed-tuple gap reflect distribution shift? Fine-tuning Wid3R on LENSCOPE would help separate the effect of the data from the architecture.
  2. The Stage-2 pose results are below Stage-1 under a different loader, whereas the final model substantially outperforms Stage-2. Which Stage-3 component drives this gain: the new scenes, losses, circular padding, longer training, or checkpoint interpolation? Does the covisibility recheck help when the scenes are held fixed? Stage-1 also gives better laser pointmap accuracy under the crop-loader evaluation; do Stages 2–3 mainly improve pose rather than geometry?
  3. Given the scale bias acknowledged in Sec. 5, could the authors report SE(3)-aligned ATE, accuracy, and completeness to evaluate the "Metric" claim in Tab. 1 directly?
1 comment
AuthorsOpenReview ↗

Response to Reviewer tEzQ

Experiment with Real heterogeneous cameras. To test independent devices we built a real heterogeneous-camera ground truth: in a laser-scanned office (nine registered stations) we calibrated and localised an Insta360 X3 (stitched panoramas), a GoPro HERO4 Session (fisheye), an iPhone 12 Pro Max, a Xiaomi 17 Pro (main and ultrawide lenses) and a Fujifilm X100VI. The median held-out bearing residual of the images is 0.07°. On mixed tuples built as in the mixed track of the paper, AUC@30 is 70.5 for MEOW, 29.9 for VGGT, 26.6 and 17.0 for the two MapAnything checkpoints and 9.2 for Wid3R, which is given the camera class. Real lenses and per-device colour pipelines are thus tested; stills do not exercise rolling shutter. On three ScanNet++ scenes with native DSLR-fisheye, iPhone and scanner-panorama images, mixed tuples give 84.4 for MEOW against 22.1 for the best baseline. In ADT (Table 12) the pinhole frames are rectified from the same Aria fisheye frames.

Q1. Single-camera tracks; Wid3R's training distribution. Yes: mixed-camera capability is a matter of training distribution, which is the thesis of the paper. With one architecture, MapAnything scores 16.4 mAA@30 on the same 2D3DS tuples and MEOW 80.4. Wid3R scores 81.9, 90.5 and 87.3 on the three single-camera laser tracks and 29.3 on the mixed track (Table 3); with a constant pinhole class in place of the true classes, its mixed score moves only to 28.3. On the single-camera tracks the adaptation costs MEOW at most 1.9 AUC@30 against its backbone (85.7 against 86.8 on pinholes, 73.2 against 75.1 on fisheyes) and adds the panorama track, 71.0 against 1.5. We have not yet fine-tuned Wid3R on LENSCOPE. Its data preprocessing is not released, so we cannot mix its original training data with LENSCOPE and may lose what it has learned; it is also trained at one 518×336 resolution and one camera type per sample, while mixed aspect ratios needed an aspect-ratio embedding in our training (Table 4).

Q2. What drives the gain. Table 8 was unclear: it showed Stage 1 under its crop loader (68.3) and the later stages under full-field-of-view resizing. Under one loader the stages improve monotonically on 2D3DS: 42.2 mAA@30 for Stage 1, 63.8 for Stage 2, 80.4 for the released weights. For Stage 3, three runs from one Stage-1 checkpoint with the same scenes, seed and 6,240 updates:

Run2D3DS mAA@30, resized / cropped inputLaser AUC@30, mixed / panorama
A: Stage-1 objective, no added component56.6 / 63.274.5 / 48.6
B: full recipe, covisibility re-check off59.2 / 62.572.8 / 44.1
C: full recipe60.9 / 62.872.5 / 47.3
  • New scenes with continued training: run A adds no component and moves 2D3DS from 42.2 to 56.6, the laser mixed track from 58.3 to 74.5.
  • Losses, embedding and circular padding, A to C: +4.3 on 2D3DS with resized input. Over a longer budget, the model without the embedding reaches 70.6† after 90 Stage-3 epochs, against 76.6† with it.
  • Covisibility re-check with scenes fixed, B to C: +1.8 on 2D3DS and +3.2 on the laser panorama track.
  • Longer training and interpolation: 76.6† after 90 epochs, 80.6† after 859; interpolation [1,2] then moves 2D3DS from 81.3 to 80.4 and the laser mixed track from 71.7 to 79.4.
  • Pose or geometry: mainly pose, as the reviewer suspects: laser mixed AUC rises from 58.3 to 79.4 from Stage 1 to the released weights. Under one loader pointmap accuracy also improves, from 0.169 m to 0.142 m.

Q3. SE(3)-aligned ATE, accuracy and completeness. The three metrics follow one rigid alignment per tuple, with no scale fitted and no ICP. The transform is fitted to the camera centres for ATE, as in the TUM RGB-D benchmark [3], and to the points for accuracy and completeness, as in VGGT [4] but with the scale fixed, fitted on half of the valid pixels and measured on the other half. Entries are MEOW / MapAnything (Apache), its starting point / MapAnything (NC, CC BY-NC v1.1), in metres; best in bold.

SE(3)-ATEAccuracyCompleteness
Laser, mixed0.77 / 2.76 / 3.150.31 / 0.56 / 0.820.25 / 0.26 / 0.26
Real cameras0.50 / 2.27 / 1.630.29 / 0.86 / 0.740.27 / 0.20 / 0.21

MEOW has the lowest trajectory error and the best accuracy on both. Completeness is level on the laser track and lower for MapAnything on the real cameras, where its accuracy error is 2.5 to 3 times that of MEOW; completeness does not penalise misplaced points. The metric claim of Table 1 thus holds with scale kept; the scale itself is 12–25% short, as Section 5 states.

† Submission ground truth; other 2D3DS values follow an axis-convention fix in the ground truth (every method +0.1 to +0.5 mAA@30, MEOW 79.9 to 80.4, no ranking change).

[1] Wortsman et al., WiSE-FT, CVPR 2022. [2] Wortsman et al., Model soups, ICML 2022. [3] Sturm et al., TUM RGB-D benchmark, IROS 2012. [4] Wang et al., VGGT, CVPR 2025.

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

MEOW adapts public MapAnything weights, using synthetic data, to jointly reconstruct pinhole, fisheye, and full-panorama views in one feed-forward pass. Its LENSCOPE engine renders procedural indoor scenes as panoramas with exact rays, depth, and poses; samples multiple camera projections; and checks covisibility after camera sampling. The reconstruction backbone is retained, with full-field-of-view resizing, a native-aspect-ratio embedding, panorama boundary handling, and adapted supervision. Inference requires neither calibration nor supplied camera-type labels. The paper reports strong zero-shot pose and point-cloud results on mixed views resampled from real panoramic captures and on a new laser-scanned office benchmark, alongside input ablations and additional evaluations.

Critical strengths
  1. A concrete, testable approach to mixed-camera training data. LENSCOPE rechecks covisibility after sampling the camera models and supplies coherent ray, depth, and pose supervision. This makes the tuples more geometrically useful than simply mixing images from different projections.
  2. Large mixed-view pose gains under a common protocol, supported by targeted controls. On heterogeneous 2D3DS tuples, MEOW reaches 79.9 mAA@30 versus 53.8 for Wid3R, which receives a camera class for every view. On the laser mixed track, it reaches 79.4 AUC@30 versus Wid3R's 29.3. Controls with public MapAnything weights, an early adapted checkpoint, resizing, and the aspect-ratio embedding help attribute the gain.
Critical weaknesses
  1. The central metric-pointmap claim is not adequately tested by the headline metrics. Trajectory error and point-cloud results use per-tuple Sim(3) alignment, which removes predicted scale error; relative-pose angular metrics also do not measure absolute scale. Appendix F reports ground-truth/predicted scale ratios of 1.14 on 2D3DS and 1.26–1.34 across the laser tracks, indicating systematic scale bias. Without unaligned 3D point, distance, or absolute pose errors, the strong results in Tables 2–3 do not establish how accurate the claimed metric outputs are (pp. 7–9, 23).
  2. Generalization to real mixed physical cameras remains unverified. Both headline mixed-camera benchmarks resample their pinhole and fisheye views from real panoramic captures. They therefore do not quantitatively test independently captured devices with real lens distortion, device-specific imaging pipelines, and stitching artifacts. The laser benchmark covers one office with 12 stations and 24 tuples per track; a physical phone/fisheye/panorama combination is shown only qualitatively without ground truth. The results support reconstruction of synthetic mixed projections of real scenes, but not yet the broader real-sensor claim (pp. 7–9, 21–23).
Questions for authors
  1. Can the authors report without Sim(3) scale alignment absolute pointmap/depth errors and camera-baseline length errors, broken down by camera type and mixed tuples, alongside comparable baseline numbers? This directly tests the metric-reconstruction claim.
  2. Is a quantitative test possible on mixed views captured by independent physical pinhole, fisheye, and panoramic devices with calibration or geometric ground truth? If not, how should the main claims be limited to resampled projections, and what evidence addresses the lens and device domain gap?
1 comment
AuthorsOpenReview ↗

Response to Reviewer fGKd

Q1 / W1. Accuracy without Sim(3) alignment, by camera type, with baselines. The table below reports results before rescaling: no scale is fitted, and each tuple is expressed in the coordinate frame of its first camera. Entries are MEOW / MapAnything (Apache) / MapAnything (NC), the only other method with metric output. Apache is the checkpoint MEOW starts from and the one in the paper's tables; NC is the stronger CC BY-NC v1.1 checkpoint. The best value is bold.

InputsBaseline length, rel. errorRange AbsRelPoint error (m)
Laser, mixed0.24 / 0.70 / 0.810.22 / 0.40 / 0.620.98 / 6.64 / 7.14
Laser, pinhole0.21 / 0.18 / 0.050.18 / 0.13 / 0.070.98 / 0.76 / 0.37
Laser, fisheye0.25 / 0.56 / 0.780.22 / 0.40 / 0.631.11 / 1.85 / 2.74
Laser, panorama0.30 / 0.95 / 0.920.24 / 1.98 / 3.442.89 / 11.5 / 15.4
2D3DS, mixed0.16 / 0.98 / 1.040.15 / 0.54 / 0.960.87 / 5.13 / 5.94
Real cameras, mixed0.24 / 1.74 / 0.990.19 / 0.39 / 0.570.76 / 3.43 / 3.08

MEOW is the most accurate on mixed, fisheye and panorama input in every column. Its point error is six to seven times lower than MapAnything's on the laser and 2D3DS mixed tuples and more than four times lower on the real cameras; MapAnything keeps its advantage on the pinhole-only tuples it was built for. The ratio of true to predicted baseline is 1.14 on 2D3DS, 1.26–1.34 on the laser tracks and 1.25 on the real cameras, consistent with Section 5. Sim(3)-aligned results remain the shape and pose protocol, as in VGGT, MonST3R, CUT3R and π3 [1–4].

Q2 / W2. Independent physical devices with geometric ground truth. We built one. We added five laser-scanned rooms, and in one of them, an office scanned from nine registered stations, we calibrated and localised five physical cameras in six imaging modes: an Insta360 X3 (stitched 360° panoramas), a GoPro HERO4 Session (action-camera fisheye), two phones of different makes (an iPhone 12 Pro Max, and a Xiaomi 17 Pro with its main and ultrawide lenses) and a Fujifilm X100VI (APS-C, fixed lens). Each mode has its own camera model: checkerboard calibration for the GoPro, the Fujifilm and both Xiaomi lenses, the capture app's corrected intrinsics for the iPhone, and an equirectangular model for the stitched panoramas. Every image is registered to the scans, as in ScanNet++ [5]; the median held-out bearing residual is 0.07°, the worst image 0.15°. This covers the three effects listed in the review: real lens distortion, device-specific imaging pipelines and stitching artifacts.

Mixed tuples are selected with the logic of the mixed track of the paper's laser benchmark, from images whose registration is reliable. Every method receives RGB only; Wid3R also receives the camera class. AUC@30 (metric of Table 3) is 70.5 for MEOW, 29.9 for VGGT, 26.6 and 17.0 for the two MapAnything checkpoints and 9.2 for Wid3R.

The laser benchmark. The reviewer notes that it covers one office. The five new rooms were scanned in four buildings. The four without real cameras extend the benchmark with the same construction: on the mixed track MEOW reaches 79.5, MapAnything (NC) 33.4 and Wid3R 21.8. All five rooms will be released with the benchmark.

Public native-camera data. On the first three scenes of MapAnything's published ScanNet++ test split, with native DSLR-fisheye, iPhone and scanner-panorama images, mixed-tuple AUC@30 is 84.4 for MEOW, 22.1 for MapAnything (NC) and 11.8 for Wid3R.

These results enter the revision as Appendix G; the claims of the paper are unchanged. We also fixed an axis convention in the ground-truth rotations of the synthesized 2D3DS views: every method gains 0.1–0.5 mAA@30, MEOW moves from 79.9 to 80.4, and no ranking changes.

[1] Wang et al., VGGT, CVPR 2025. [2] Zhang et al., MonST3R, ICLR 2025. [3] Wang et al., CUT3R, CVPR 2025. [4] Wang et al., π3, ICLR 2026. [5] Yeshwanth et al., ScanNet++, ICCV 2023.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers for their careful reading. The replies keep the claims of the paper and support them with the experiments the reviews ask for. We also clarify points that were not explained clearly enough in the submission.

  1. Five new rooms in four buildings, all laser-scanned (fGKd, tEzQ, p5zb). In one of them we calibrated and localised five physical cameras in six imaging modes against the scans, which gives a complete real heterogeneous-camera ground truth. On its mixed tuples, built as in the mixed track of the paper, MEOW reaches 70.5 AUC@30 against 29.9 for the best baseline. The other four rooms extend the laser benchmark: 79.5 against 33.4 on the mixed track. All five rooms will be released with the benchmark.
  2. Accuracy at metric scale, without Sim(3) alignment (fGKd, tEzQ). On the laser mixed tuples MEOW's point error is 0.98 m, against 6.64 m and 7.14 m for the two MapAnything checkpoints, the only other method with metric output; on the real cameras, 0.76 m against 3.43 m and 3.08 m.
  3. Training at equal budget (p5zb, tEzQ). With the architecture unchanged, 2D3DS mAA@30 rises from 15.4 for the public MapAnything checkpoint to 42.2 after Stage 1, 56.6 with continued adaptation on LENSCOPE alone and 60.9 with the full recipe at the same budget. Over a longer budget, the model without the aspect-ratio embedding reaches 70.6 after 90 Stage-3 epochs, against 76.6 with it (both under the submission's ground truth).
  4. Baselines (p5zb). On identical image pairs MEOW is ahead of the released CAM3R checkpoint on all three test sets (77.5 against 17.8 mAA@30 on the laser benchmark). A perspective-conversion pipeline with ground-truth calibration is 0.16 m more accurate on the laser office, five times slower, and fails on part of 2D3DS (mean point error 8.46 m against 0.87 m).
  5. More real cameras, on public data. ScanNet++ adds further real cameras with laser ground truth: a DSLR with a fisheye lens, an iPhone and the scanner's panoramas. On three of its scenes, mixed tuples of the three give 84.4 AUC@30 for MEOW against 22.1 for the best baseline. Together with item 1, this is a quantitative test of generalisation to real heterogeneous cameras, built as reliably as we could.

Revision. The revision retains the paper's original claims and overall structure. The additional experiments are reported in a new Appendix G, which also shows the reconstructed point clouds beside the laser reference, and the relevant passages have been revised to clarify points that were insufficiently explained in the original submission. The comparison with CAM3R, whose checkpoint was released after the submission, is given in our reply to Reviewer p5zb. We also correct the axis convention used for the ground-truth rotations of the synthesized 2D3DS views. This correction increases mAA@30 by 0.1–0.5 for all methods, including from 79.9 to 80.4 for MEOW, without changing any method ranking. In addition, we remove one duplicated panorama tuple from Table 9, leaving 19 tuples, and revise several implementation details in Section 3 and Appendices A, B, and F to ensure consistency with the training code. We have also updated the supplementary material with an offline viewer of the additional results. We invite the reviewers to consult the revised submission for the complete results and corresponding changes.

Photos, results and interactive point clouds are on an anonymous page; the link is in a separate comment to the reviewers and area chairs. The page carries no analytics, cookies or external requests.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.