acceptodds
Under review as a conference paper at ICLR 2027

Slot3R: Set-Associative Spatial Memory for Streaming 3D Reconstruction

Abstract

Streaming 3D reconstruction must preserve evidence from each frame while processing an expanding scene online. Spatial memory is a natural fit because it organizes history by reconstructed 3D location. Yet Point3R uses spatial proximity both to associate a new observation with an existing memory entry and to decide whether to fuse it, conflating co-location with state identity. Because pointers summarize image patches, nearby pointers may encode distinct surfaces, viewpoints, or visibility conditions; averaging them can destroy complementary evidence before later frames disambiguate it. We argue that location should determine address, not whether observations must merge. Slot3R realizes this principle as a training-free, set-associative retrofit that lets multiple states coexist at a shared address while keeping the pretrained Point3R backbone frozen. A bounded sparse readout further decouples persistent storage from per-frame decoder access. At 300–500 sampled frames, Slot3R reduces Point3R’s point-cloud accuracy error (Acc) by 57.1%–63.1% on 7Scenes and 64.0%–72.0% on NeuralRGBD, lowers Sim(3)-aligned absolute trajectory error (ATE) on all three pose benchmarks, and remains competitive on video-depth estimation. It completes all evaluated settings from 600 to 1000 sampled frames at about 19 FPS under the same protocol, whereas Point3R and InfiniteVGGT run out of memory at 800 frames and beyond.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction

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 3 reviews, above 85% of the others.
Reviewer_Bf3yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Slot3R, a training-free modification of Point3R for streaming 3D reconstruction. It replaces location-only fusion with a set-associative spatial memory that can retain multiple feature states at one spatial address, and employs a bounded sparse readout of at most 640 memory tokens per frame. With a frozen Point3R backbone, the method substantially improves 3D reconstruction accuracy over Point3R on 7Scenes and NRGBD at 300–500 sampled frames. It also completes the reported 600–1000-frame tests, whereas Point3R runs out of memory at 800 frames and beyond.

Critical strengths
  1. The paper identifies a reasonable limitation in Point3R and proposes a simple, well-motivated modification to address it. The substantial improvements in reconstruction accuracy across both datasets demonstrate the effectiveness of the proposed approach without additional training.

  2. The paper is clearly written and well organized. The experimental evaluation is fairly comprehensive, with extensive ablation studies and runtime analyses that provide useful support for the proposed design choices.

Critical weaknesses
  1. The claimed motivation is not sufficiently validated. While the K=1 to K=8 ablation shows that increasing the number of slots improves reconstruction accuracy, the paper does not directly demonstrate that Point3R incorrectly merges spatially nearby but distinct feature states, or that Slot3R preserves such complementary information. Although Table 4 provides a useful component ablation, the K-way memory setting changes addressing and updating together, so it remains unclear how much of the improvement can be attributed specifically to avoiding premature fusion. Qualitative analysis of incorrectly merged states in Point3R and the corresponding slots in Slot3R, together with a controlled ablation isolating the fusion strategy, would better support the central claim.

  2. The scalability and retrieval robustness of Slot3R under broader scene exploration remain insufficiently evaluated. The local readout relies on the previous frame's same-index patch locations as spatial priors (Sec. 3.4), which may be unreliable under large camera motion or sparse temporal sampling. Although the paper includes a k=20 experiment, it does not directly analyze the robustness of the retrieval strategy under such conditions. Moreover, the long-sequence experiments mainly evaluate increasing frame counts rather than expanding scene coverage. While the sparse readout bounds the number of tokens accessed by the decoder, persistent storage can still grow with the number of occupied buckets, up to the fixed grid capacity. Experiments on larger SLAM-style scenes, along with memory usage and runtime analysis as scene coverage expands, would provide stronger evidence for the scalability of the approach.

Questions for authors
  1. The qualitative results show that Slot3R produces clearer reconstruction boundaries than Point3R, but it remains unclear whether different slots actually capture distinct surface or viewpoint states, or whether the improvement mainly comes from greater capacity within each spatial bucket. Could the authors visualize the image patches or 3D regions associated with different slots within the same bucket, particularly near object boundaries? Additionally, a comparison that holds the total storage budget and readout fixed while varying the fusion strategy, together with the corresponding storage usage and reconstruction errors, would help clarify the role of multi-slot storage.

  2. How reliable is the previous-frame patch prior under large camera motions, occlusions, or temporally sparse inputs? Could the authors provide an analysis of local retrieval hit rates and reconstruction errors under these conditions, and quantify how much global anchors contribute when local retrieval becomes unreliable? The current-frame-prior ablation in Table 5 does not fully address these challenging scenarios.

  3. How do memory usage and inference latency change as the camera continuously explores new areas, such as moving through an increasing number of previously unseen rooms and corridors, rather than repeatedly observing the same region? Could the authors report occupied buckets, stored slots, peak GPU/CPU memory, and per-stage latency as the explored area grows, with comparisons to Point3R ?This would help distinguish bounded readout costs from the growth of persistent memory.

  4. For the Cartesian baseline in Table 14, what are the voxel size, spatial range, effective bucket count, and occupancy statistics? Since different spatial discretizations may result in substantially different storage capacities and resolutions, could the authors provide a comparison under matched effective capacity and comparable spatial resolution? This would help clarify whether the improvement comes from spherical hashing itself or from differences in memory allocation.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Bf3y

We thank the reviewer for the detailed and constructive feedback and for recognizing the motivation and reconstruction improvements. We appreciate the opportunity to clarify the memory mechanism and its behavior across the evaluated settings. We have added experiments and analyses to address the concerns about the memory mechanism, retrieval robustness, storage usage, and hashing comparison. We have also uploaded a revised manuscript for the reviewer’s consideration.

  1. Evidence for the memory mechanism

We agree with the reviewer that the original submission would benefit from more direct evidence. Building on the compact illustration in Fig. 2(b), we have added more extensive visualizations in Fig. 3, comparing corresponding local regions in Point3R and Slot3R and identifying slots within a shared bucket. These examples show different local appearances associated with separately stored states at the same spatial address. This supports our central motivation: sharing a spatial address does not imply appearance identity, and multiple slots allow these different states to coexist. The visualizations show final stored states rather than traces of individual Point3R fusion events.

As a complementary controlled test, we also added a random-fusion control (Table 6; Appendix A.5), replacing similarity-based fusion-partner selection with random selection while keeping the backbone, confidence filtering, spatial grid, and sparse readout unchanged. At nearly identical average final memory sizes, Core (K=8K=8) achieves an Acc of 0.0408, compared with 0.0827 for random fusion (K=5K=5), using 754.6 and 753.6 entries, respectively. At the same capacity K=8K=8, random fusion still performs worse despite storing 58.3% more entries. These results support the benefit of feature-guided fusion beyond simply retaining more tokens.

Fusion strategy Acc ↓  Comp ↓  NC ↑ Avg. final stored entries
Point3R0.11320.03770.6116890.2
Core Slot3R (K=8K=8)0.04080.01510.6662754.6
Random fusion (K=5K=5)0.08270.026830.61702753.6
Random fusion (K=6K=6)0.076540.02180.61821880.8
Random fusion (K=8K=8)0.075520.018560.618171194.1

Nine NeuralRGBD sequences, k=2k=2, up to 300 sampled frames. An incoming state uses a free slot when available; when fusion is needed, the random control selects a partner randomly among occupied slots in the same bucket. Further details are provided in Table 6 and Appendix A.5.

  1. Previous-frame prior robustness and the role of global anchors

We thank the reviewer for encouraging us to examine the previous-frame prior more directly. We added an experiment in Appendix A.7, Tables 15–16 to examine how increasing camera motion affects the reliability of the previous-frame spatial prior, and how much global anchors can mitigate the resulting impact on reconstruction. We vary the sampling interval k∈{2,10,20,30,50}k\in\{2,10,20,30,50\} to obtain increasingly large inter-frame motions. We assess prior reliability through neighborhood-index recall, using current predicted positions as a post-hoc reference. To evaluate the contribution of global anchors, we compare 0 versus 128 anchors on identical input frames at each interval, keeping the total readout cap fixed at 640 tokens.

NeuralRGBD

kkRot. (°)Trans. (m)Recall (%)Acc, G=0 ↓Acc, G=128 ↓Acc gain (%)
20.750.02997.470.042980.04017+6.55
103.700.14490.070.042630.04065+4.63
207.240.28882.890.043820.04173+4.76
3010.960.43076.210.048680.04388+9.86
5017.420.70565.880.050580.04319+14.61

7Scenes

kkRot. (°)Trans. (m)Recall (%)Acc, G=0 ↓Acc, G=128 ↓Acc gain (%)
21.720.02194.870.026610.02363+11.21
105.680.10082.350.029380.02633+10.38
2010.410.19671.630.033380.02939+11.95
3014.520.28463.420.036190.03155+12.81
5022.420.44350.660.042720.03671+14.08

G is the number of global anchors. Motion uses ground-truth poses; recall is averaged over the two configurations. Acc gain is the relative error reduction from G=0 to G=128.

The results show that larger inter-frame motion is accompanied by lower neighborhood-index recall, indicating that the previous-frame prior retrieves less of the reference neighborhood. Global anchors generally provide larger reconstruction benefits under these less reliable retrieval conditions. At k=50k=50, adding anchors reduces Acc error relative to the no-anchor variant by 14.61% on NeuralRGBD and 14.08% on 7Scenes. Although the gains are not strictly monotonic across intervals, the overall trend supports global anchors as a fallback source of nonlocal evidence that helps compensate for less reliable local retrieval.

AuthorsOpenReview ↗

Response to Reviewer Bf3y

  1. Persistent-memory usage and bucket occupancy

We appreciate the reviewer’s suggestion to examine storage and latency as more of the scene is observed. We added sequence-averaged memory statistics in Appendix A.6, Table 14 and Fig. 5, comparing Point3R’s stored pointers with Slot3R’s stored slots over 200–500 sampled frames. Slot3R stores fewer entries in seven of eight dataset–length settings; the exception is NeuralRGBD at 500 frames, where it stores 11.2% more. Across these settings, only 3.08%–6.02% of the 4096 addresses are occupied, while occupied buckets use 53.55%–72.27% of their slot capacity. These statistics show that sparse spatial allocation coexists with substantial use of multiple slots within occupied regions.

To examine memory growth within a sequence, we additionally profile seven NeuralRGBD test sequences with 500 sampled frames in Appendix A.13, Table 24. We measure cumulative observed surface coverage from ground-truth depth and poses, independently of either model. From frames 1–50 to the final 50-frame block, mean coverage rises from 41.7% to 100% of each sequence’s final observed surface; Slot3R’s occupied buckets grow from 102.3 to 272.0 and stored slots from 564.1 to 1081.1, while Point3R’s stored entries grow from 798.1 to 934.6. As storage grows, Slot3R’s selection-plus-decoder latency remains near 21 ms/frame (19.98–21.34), whereas Point3R’s decoder latency increases from 24.51 to 28.32 ms/frame. The final GPU peaks are comparable, with Slot3R slightly higher (12.68 versus 12.52 GiB). This supports the distinction between growing persistent storage and bounded decoder access under the evaluated sequences.

FramesGT cumulative surface coverage (%)Slot3R occupied bucketsStored states: Slot3R / Point3RGPU peak: Slot3R / Point3R (GiB)Decoder path: Slot3R / Point3R (ms/frame)
1–5041.7102.3564.1 / 798.110.09 / 10.0919.98 / 24.51
201–25067.5145.0724.7 / 841.611.06 / 10.9720.60 / 26.22
451–500100.0272.01081.1 / 934.612.68 / 12.5221.22 / 28.32

Averages over seven NeuralRGBD sequences with 500 sampled frames. Stored states are slots for Slot3R and pointers for Point3R; Point3R has no buckets. Slot3R's decoder path includes selection and decoder forwarding, whereas Point3R's includes decoder forwarding. We kindly refer the reviewer to Appendix A.13, Table 24 for all 50-frame intervals.

  1. Spherical versus voxel hashing under comparable storage

We thank the reviewer for identifying an important limitation in our original comparison: the voxel variant stored fewer slots, making it difficult to attribute the performance difference solely to the hash design. We reran the spherical-versus-voxel comparison with closely matched average final stored-slot counts (Appendix A.9, Table 19). Spherical and voxel hashing store 1160.86 versus 1142.71 slots on NeuralRGBD and 1197.11 versus 1182.78 on 7Scenes. Under these comparable storage levels, spherical hashing achieves lower Acc on both datasets (0.048954 vs. 0.052406 and 0.029199 vs. 0.034709, respectively), while also improving Comp and NC. These results support the benefit of spherical addressing beyond differences in average stored-memory size. For the full configurations and occupancy statistics, we kindly refer the reviewer to Appendix A.9.

Dear Reviewer Bf3y,

We would like to kindly follow up on our rebuttal and revised manuscript. To address your main concerns, we have added visualizations and a random-fusion control to further validate our memory mechanism (Fig. 3, Table 6), analyzed the robustness of the previous-frame prior and the role of global anchors under increasing camera motion (Appendix A.7, Tables 15–16), reported memory occupancy and per-stage latency as scene coverage grows (Appendix A.6 and A.13), and rerun the spherical-versus-voxel comparison under comparable storage (Appendix A.9, Table 19).

We would greatly appreciate it if you could revisit these additional results and let us know whether any concerns remain before the discussion period ends.

Thank you again for your time and constructive feedback!

Best regards,

The Authors

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

This paper addresses the potential information loss caused by merging memory pointers solely based on spatial proximity in Point3R and proposes a training-free Set-Associative Spatial Memory framework. The authors argue that spatial proximity does not necessarily imply state identity. Accordingly, the proposed K-Way Memory allows multiple distinct feature states to coexist within the same spatial region, while feature similarity and confidence guide memory merging, insertion, and replacement. Additionally, a Sparse Spatial Readout mechanism combines local spatial retrieval with global anchors to limit the number of memory tokens. Experimental results shows that the proposed method improves long-sequence 3D reconstruction accuracy, while maintaining inference efficiency.

Critical strengths
  1. Meaningful research problem and clear methodology. This paper addresses the potential issue of inappropriate memory fusion in streaming 3D reconstruction. The motivation is well established, the overall approach is reasonable. Moreover, its training-free nature enhances its practical applicability, making this a meaningful contribution.
  2. Significant performance improvements with comprehensive experiments. The proposed method shows accuracy improvements across datasets, while extensive ablation studies further evaluate its effectiveness.
Critical weaknesses
  1. Insufficient robustness analysis of the key merging threshold. Although the supplementary material provides sensitivity analyses for several hyperparameters, the key merging threshold τmerge\tau_{\text{merge}} is not investigated. Since this threshold directly determines whether different memory states are merged, further justification of its selection and sensitivity analysis would be valuable.
  2. Limited analysis and interpretation of the experimental results.

I. Although the method achieves improvements across multiple datasets, the experimental analysis mainly focuses on reporting quantitative results, with limited discussion of the underlying reasons for these performance changes.

II. The paper lacks visualizations or quantitative analyses that directly illustrate how Set-Associative Spatial Memory operates and contributes to the observed improvements.

III. The performance degradation on certain pose estimation and metric-scale depth metrics is not sufficiently explained from the perspective of the proposed method (RPE rot in camera pose estimation and AbsRel in metric-scale depth estimation) .

Questions for authors
  1. Could the authors provide sensitivity analysis results for the key merging threshold τmerge\tau_{\text{merge}} to further demonstrate the robustness of the proposed method? Q2. Could the authors provide further quantitative analyses to explain the underlying reasons for the performance improvements brought by Set-Associative Spatial Memory, as well as discuss why the method leads to performance degradation on certain pose estimation and metric-scale depth metrics?
4 comments
AuthorsOpenReview ↗

Response to Reviewer fJqY

We thank the reviewer for the constructive comments and for recognizing the clear motivation and strong empirical results. We appreciate these suggestions, which have helped us strengthen the experimental analysis. We have added further experiments and analyses to address the two main concerns below and have also uploaded a revised manuscript for the reviewer’s consideration.

  1. Sensitivity to the merging threshold

To address the reviewer's suggestion, we have added a sensitivity analysis in Appendix A.3, Table 9, varying τmerge\tau_{\mathrm{merge}} from 0.50 to 0.95 on 7Scenes and NeuralRGBD while keeping all other settings fixed. The best results across the reported metrics fall within 0.70–0.90: 0.70 gives the lowest Acc and Comp on 7Scenes, while 0.90 gives the lowest on NeuralRGBD. These results support this interval as a reasonable operating range under our evaluation; we retain 0.90 as a common threshold across datasets.

τmerge\tau_{\mathrm{merge}}7Scenes Acc ↓7Scenes Comp ↓7Scenes NC ↑NeuralRGBD Acc ↓NeuralRGBD Comp ↓NeuralRGBD NC ↑
0.500.027320.017140.584820.041110.017940.67692
0.600.025930.016580.584910.041060.017720.67609
0.700.024860.016180.585490.041170.018050.67099
0.800.025510.016750.586210.039570.018310.67920
0.900.025850.017350.585390.038530.017370.67770
0.950.026550.016960.584980.038890.017820.67200

All other settings are fixed; both datasets use 200 sampled frames at k=2k=2. Full results are in Appendix A.3, Table 9.

  1. Analysis of the results and the memory mechanism

We agree with the reviewer that the original submission would benefit from more direct evidence and interpretation, and we appreciate this suggestion. We have added visualizations in Fig. 3, comparing corresponding local regions in Point3R and Slot3R and identifying slots within a shared bucket. These examples illustrate how separately stored states retain different local appearances despite sharing a spatial address. This supports our central motivation: sharing a spatial address does not imply appearance identity, and multi-slot storage allows these different states to coexist rather than forcing them to merge solely by location.

To further address this concern, we also added a random-fusion control (Table 6; Appendix A.5) in which fusion partners are selected randomly from occupied slots within the same bucket rather than based on feature similarity. At nearly identical average final memory sizes, Core (K=8K=8) achieves an Acc of 0.0408, compared with 0.0827 for random fusion (K=5K=5), using 754.6 and 753.6 entries, respectively. Even at K=8K=8, random fusion performs worse despite storing more entries. These results support the benefit of decoupling spatial addressing from fusion decisions: using feature similarity to select fusion partners improves reconstruction beyond what can be explained by retaining more tokens alone.

Fusion strategy Acc ↓  Comp ↓  NC ↑ Avg. final stored entries
Point3R0.11320.03770.6116890.2
Core Slot3R (K=8K=8)0.04080.01510.6662754.6
Random fusion (K=5K=5)0.08270.026830.61702753.6
Random fusion (K=6K=6)0.076540.02180.61821880.8
Random fusion (K=8K=8)0.075520.018560.618171194.1

Nine NeuralRGBD sequences, k=2k=2, up to 300 sampled frames. An incoming state uses a free slot when available; when fusion is needed, the random control selects a partner randomly among occupied slots in the same bucket. Further details are provided in Table 6 and Appendix A.5.

We also appreciate the reviewer’s attention to the pose and depth results. To investigate the pose and depth degradation, we added a comparison between Core Slot3R with sparse readout and the same model with sparse readout disabled, allowing the decoder to access all stored memory tokens (Appendix A.11, Tables 21–22). Removing sparse readout improves translation and rotation RPE on all three pose benchmarks; for example, on ScanNet, translation RPE decreases from 0.0322 to 0.0288 and rotation RPE from 1.4235 to 1.1090. Together with the FPS improvement from sparse readout in our separate runtime comparison (16.658 to 18.269 FPS; Table 20), this indicates that Core’s relative-pose degradation partly reflects a trade-off between pose accuracy and inference efficiency. In contrast, depth performance changes little when sparse readout is removed: on KITTI, metric-scale AbsRel is 0.1790 with sparse readout and 0.1795 with full readout, compared with 0.1690 for Point3R. We acknowledge the remaining metric-scale depth degradation on Bonn and KITTI relative to Point3R as a limitation of our method. However, under per-sequence alignment, both Core variants achieve lower AbsRel than Point3R on all three depth datasets.

Dear Reviewer fJqY,

We would like to kindly follow up on our rebuttal and revised manuscript. In response to your main concerns, we have added a merging-threshold sensitivity analysis (Appendix A.3, Table 9), qualitative visualizations and a controlled random-fusion experiment (Fig. 3, Table 6), as well as sparse-versus-full-readout comparisons to investigate the pose and depth degradation (Appendix A.11, Tables 21–22).

We would greatly appreciate it if you could revisit these additional results and let us know whether any concerns remain before the discussion period ends.

Thank you again for your time and constructive feedback!

Best regards,

The Authors

Thank you for the clarifications and additional experiments. My concerns are resolved and I will keep my rating.

Thank you very much for taking the time to review our response and additional experiments. We sincerely appreciate your constructive feedback and are glad that our clarifications have addressed your concerns. Your suggestions have helped us strengthen the experimental analysis and improve the clarity of our paper.

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

This paper introduces a training-free streaming 3D reconstruction method based on Point3R. It is a training-free method. Previous Point3R merged the memory based on spatial distance, which may lose some important information because the close position may not mean the similar observation. This paper try to solve this issue by proposing a sphere hash assigns observations to spatial buckets, each containing multiple slots. The feature cosine similarity and confidence will be used to merge or replaced, or rejected the observations. Experiments shows the superior performance comparing to the baseline Point3R and other streaming based 3D reconstruction methods.

Critical strengths
  1. The motivation is clear. Close position features should not be simply merged because different-angle observations contains different information.
  2. Training free and large improvements in terms of reconstruation quality and pose estimation.
Critical weaknesses
  1. I have concerns about the design of the sphere hash. When the camera moves in the large scenes (such as kitti dataset), the bucket in the farther place will be very sparse, which will affect the accuracy of the memory.
  2. Large scene reconstruction performance should be further explored. Current experiments only reports the video depth estimation performance and is slightly worse than the based Point3R in metric-scale.
Questions for authors
  1. Does the sphere has the fixed origin? Will the sphere move with the camera?
  2. Will the feature outside the sphere max distance be stored?
3 comments
AuthorsOpenReview ↗

Response to Reviewer yKih

We thank the reviewer for the constructive comments and questions, and for recognizing the clear motivation of our work. We appreciate the opportunity to clarify the addressing scheme and the scope of our large-scale evaluation. Below, we clarify the spherical addressing scheme and discuss the related concerns about large-scale exploration. We have also uploaded a revised manuscript for the reviewer’s consideration.

We would like to respectfully clarify that the spherical grid is centered at the camera center of the first input frame—the origin of Point3R’s global coordinate system—and remains fixed as the camera moves. Candidates beyond ρmax⁡\rho_{\max} are mapped to the outermost radial bin with their angular indices retained rather than discarded. As the reviewer notes, our KITTI evaluation currently focuses on video depth (Table 2). Core Slot3R achieves lower per-sequence aligned AbsRel than Point3R, although its direct metric-scale AbsRel is slightly higher.

We recognize that these questions reflect a broader concern about the robustness of our method in large-scale scenes. In response to this helpful comment, we added an adaptive-center module triggered when the proportion of out-of-range write pointers exceeds a threshold, with a center shift accepted only if it reduces overflow (Appendix A.12, Table 23). We checked its activation across point-cloud reconstruction (NeuralRGBD and 7Scenes), pose estimation (ScanNet, Sintel, and TUM-Dynamic), and depth estimation (Bonn and KITTI). Under the tested protocols, no center updates occur on either point-cloud dataset, any of the three pose datasets, or Bonn depth; in these settings, the variant falls back to fixed-origin Core. KITTI depth is the only tested setting that activates center adaptation. On KITTI, the 5% and 10% overflow thresholds reduce metric-scale AbsRel by 4.59% and 4.07%, respectively, relative to fixed-origin Core. However, both variants slightly worsen per-sequence aligned depth, indicating a modest, metric-dependent benefit rather than a uniform improvement. These mixed results show that the tested center-adaptation strategy alone is not a complete remedy. We therefore retain fixed-origin Core as the default and report center adaptation as an optional extension. For the complete evaluation, we kindly refer the reviewer to Appendix A.12, Table 23.

We agree that large-scale global reconstruction remains a limitation of our current method, and we thank the reviewer for highlighting it. We previously generated and visually inspected point-cloud reconstructions for large-scene sequences and observed geometric distortion in both Point3R and Slot3R. One possible explanation is that pose and geometry estimation errors accumulate over long or complex camera trajectories, making it difficult to maintain a globally consistent reconstruction. Since Slot3R retains Point3R’s frozen backbone and prediction heads, some of these difficulties may be shared with the underlying prediction pipeline.

Thank authors for the clarification and the additional adaptive-center experiments. I will maintain my original rating

Thank you for taking the time to review our response and for acknowledging our work. We appreciate your constructive feedback, which helped us clarify the method and strengthen the experimental evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.