- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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.
-
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.
-
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
Response to Reviewer Bf3y
- 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.
| Frames | GT cumulative surface coverage (%) | Slot3R occupied buckets | Stored states: Slot3R / Point3R | GPU peak: Slot3R / Point3R (GiB) | Decoder path: Slot3R / Point3R (ms/frame) |
|---|---|---|---|---|---|
| 1–50 | 41.7 | 102.3 | 564.1 / 798.1 | 10.09 / 10.09 | 19.98 / 24.51 |
| 201–250 | 67.5 | 145.0 | 724.7 / 841.6 | 11.06 / 10.97 | 20.60 / 26.22 |
| 451–500 | 100.0 | 272.0 | 1081.1 / 934.6 | 12.68 / 12.52 | 21.22 / 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.
- 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
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.
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=8) achieves an Acc of 0.0408, compared with 0.0827 for random fusion (K=5), using 754.6 and 753.6 entries, respectively. At the same capacity K=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.
Nine NeuralRGBD sequences, k=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 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} 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
7Scenes
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=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.