acceptodds
Under review as a conference paper at ICLR 2027

MixPT: Point Transformer with Interleaved Window and Global Attention

Abstract

Point cloud transformers aim to capture detailed local geometry and relationships across a scene at a manageable computational cost. Two common approaches are window attention and global attention. They differ in how they enable scene-wide interaction, which refers to information exchange among tokens across a scene. Window attention reduces computational cost by restricting self-attention to groups of tokens called windows; in serialized designs, each window contains a fixed number of consecutive tokens from a spatially ordered sequence. Because each block only directly connects tokens within the same window, information exchange across windows relies on changes in window membership across blocks and hierarchical pooling. Global attention, in contrast, directly connects all tokens within a scene, but computing attention over all token pairs can be costly when the number of tokens is large. To balance computational cost and scene-wide interaction, we introduce MixPT, a point transformer that uses a mix of window and global attention blocks in an interleaved schedule, while retaining a single fixed serialization pattern for its window blocks. For scenes with a large number of tokens, MixPT applies global attention to spatially pooled tokens and returns residual updates to the original-resolution stream. Across indoor and outdoor benchmarks, MixPT achieves competitive accuracy with lower latency than representative window and global attention baselines. We further demonstrate the feasibility of scaling the attention stage in depth by repeating the interleaved pattern of window and global attention blocks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point Transformer, Point Cloud, 3D Scene Understanding, 3D Vision

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_AVNPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MixPT reuses a fixed serialization order within a single attention stage and interleaves window and global attention, for example with a WWGWWGWW schedule. For large scenes, the global branch pools spatial tokens and broadcasts residual updates back to the original resolution. Evaluation covers four semantic segmentation datasets and two instance segmentation datasets. Its main contribution is a practical trade-off between accuracy and single-forward latency.

Critical strengths
  • The design is simple and addresses a concrete computational bottleneck. Fixed serialization and occasional global blocks reduce repeated token reordering. Serialization-related latency on nuScenes decreases from 10.61 ms for LitePT-S to 1.01 ms.

  • MixPT-S offers a useful accuracy–latency trade-off across four semantic segmentation datasets, reducing single-forward latency by approximately 19%–32% over LitePT-S while improving no-TTA mIoU by 0.2–0.5 percentage points.

  • Controlled ablations support combining fixed windows with global attention. Adding two global blocks increases ScanNet mIoU from 73.7 to 76.0, with latency increasing only from 17.16 to 17.48 ms.

Critical weaknesses
  • Conceptual novelty is limited, and the contributions of individual changes are not fully isolated. The paper acknowledges prior local/global attention designs, while also changing stage structure, pooling/upsampling, and normalization relative to LitePT without a complete step-by-step ablation

  • Accuracy gains are small and not consistent across tasks. Semantic segmentation improvements over LitePT-S are only 0.2–0.5 points without multiseed variability, while ScanNet instance segmentation mAP is lower than LitePT-S* (41.1 versus 41.7).

Questions for authors

see Weaknesses

============================================

Thanks for authors' responses. After reading other reviewers' comments, I am willing to keep my initial score.

3 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for your valuable feedback. It gave us a chance to think more carefully about our contributions.

Regarding conceptual novelty. We agree that prior work has explored using both local and global attention. We think our contribution is a different way to address the computational challenges of attention in scene-level point clouds. Our analysis examines how effectively changing window layouts enables information exchange and whether global attention is needed at every block. This led us to improve computational efficiency in point transformers.

Regarding the contributions of individual changes. We also evaluated the intermediate steps from LitePT-S to MixPT-S using three random seeds to see how the changes affect accuracy and latency (Table R3-1).

Table R3-1: Incremental comparison (w/o TTA, mIoU mean ± stdev.)

StepChange from previous stepScanNet #ParamsScanNet mIoU (%)ScanNet Latency (ms)SemanticKITTI #ParamsSemanticKITTI mIoU (%)SemanticKITTI Latency (ms)
0LitePT-S: two attention stages: 6 window blocks (width 252) → spatial pooling → 2 window blocks (width 504).12.7M75.2 ± 0.2925.712.7M69.7 ± 0.1736.4
1Merge attention stages: 8 window blocks at the first stage's resolution and width (252); remove the corresponding decoder level.7.8M74.6 ± 0.1921.17.8M69.3 ± 0.0631.0
2Widen the merged stage: width 252→352, heads 14→16, and RoPE (6,6,6)→(8,8,6).13.7M74.9 ± 0.2921.613.6M69.7 ± 0.1233.5
3Remove block-entry LayerNorm; retain attention/MLP pre-norm.13.6M74.9 ± 0.3421.313.6M69.2 ± 0.6833.0
4Fix the serialization order: four shuffled orders → one fixed order (TZ), with layout reuse across all eight blocks.13.6M73.9 ± 0.1517.213.6M68.5 ± 0.0927.9
5Replace two window blocks with global blocks: WWGWWGWW (MixPT-S).13.6M75.9 ± 0.3117.514.1M69.9 ± 0.1428.2
AuthorsOpenReview ↗

Response to Weakness 2

Regarding accuracy improvements. Our main focus was to improve efficiency while keeping accuracy competitive. We can then use the time saved to run larger models for higher accuracy. For example, MixPT-L achieves higher accuracy on ScanNet than PTv3 (76.7 vs. 76.5 mIoU) with lower latency (33.4 vs. 48.1 ms) as shown in Table R3-2.

Table R3-2: Accuracy-latency comparison (w/o TTA)

MethodSemanticKITTI mIoU (%)SemanticKITTI Latency (ms)ScanNet mIoU (%)ScanNet Latency (ms)
PTv369.663.676.548.1
LitePT-S69.836.475.525.7
Volt-S69.4140.376.423.0
MixPT-S (Ours)70.028.276.017.5
MixPT-B (Ours)70.456.476.430.2
MixPT-L (Ours)70.672.476.733.4

Dear Reviewer,

Thank you for your helpful feedback. We appreciate your careful consideration of our responses. We have also added a step-by-step ablation from LitePT-S to MixPT-S to the revised manuscript (Table A9, p. 21).

We would be happy to answer any remaining questions during the discussion period. Thank you for your time.

Reviewer_26z9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MixPT proposes a point-cloud backbone that interleaves fixed-order serialized window attention with global attention. It retains LitePT's high-resolution sparse-convolution stages and part of its decoder, while consolidating multiple attention stages into one. For scenes with many tokens, the global modules operate on spatially pooled tokens and map residual updates back to the original-resolution feature stream. Experiments cover semantic segmentation on SemanticKITTI, Waymo, nuScenes, and ScanNet, as well as instance segmentation on ScanNet and ScanNet200. The method achieves competitive accuracy and lower single-forward latency than representative window-attention and global-attention baselines.

Critical strengths
  1. The efficiency improvements are consistent across different settings. Tables 1–2 show that MixPT-S reduces latency relative to LitePT-S from 36.4, 53.0, 27.8, and 25.7 ms to 28.2, 42.8, 19.2, and 17.5 ms on SemanticKITTI, Waymo, nuScenes, and ScanNet, respectively, corresponding to reductions of approximately 19–32%. Meanwhile, mIoU without TTA improves by 0.2, 0.4, 0.3, and 0.5 percentage points, respectively. Results on both indoor RGB-D scenes and outdoor LiDAR scenes support the method's practical value as an efficient backbone.

  2. The key design is supported by relevant experiments. Table 4 holds the depth and width of the attention stage fixed. Replacing two of the eight fixed-window blocks with global blocks improves mIoU from 68.6 to 70.0 on SemanticKITTI and from 73.7 to 76.0 on ScanNet, while increasing full-model latency by only approximately 0.32 ms in each case. The ScanNet comparison without global pooling also maintains the same parameter count.

  3. Implementation, evaluation, and applicability are documented in sufficient detail. The appendix provides model configurations, training recipes, EMA settings, baseline reproduction strategies, TTA, fragment aggregation, and synchronized timing details. The paper explicitly distinguishes single-forward latency from accuracy evaluation that aggregates multiple fragments.

Critical weaknesses
  1. The novelty primarily lies in a specific integration of existing designs, and attribution of the gains over LitePT remains incomplete. The high-resolution convolutional stages, retained decoder components, and axial PointROPE follow LitePT [1], while serialized windows have precedents such as PTv3 [2]. Section 2 explicitly acknowledges prior uses of interleaved local and global attention, and asymmetric 3D RoPE allocation also has precedents, including Volt [3]. Thus, the identifiable new contribution is primarily the combination of fixed serialization-layout reuse, a single attention stage, periodic global interaction, and a pooled residual branch, together with an evaluation of its efficiency. This raises concerns about limited novelty.

  2. Comparisons with related hybrid interaction methods are insufficient, and the source of each baseline result should be identified. PTv3, LitePT-S, and Volt-S cover representative window-attention and global-attention approaches, but do not adequately cover designs that jointly process local and global information. I recommend prioritizing a comparison with SP²T [4], which uses proxy representations and dual-stream local/global interaction and is directly relevant to the research question of this paper. If the authors wish to further establish advantages over other efficient backbones, KPConvX [5] could be included on ScanNet as an additional modern geometric convolution baseline. The main tables also mix results from different sources; each entry should identify its source, and the authors should clarify whether the timing configuration in Table 1 matches the corresponding accuracy configuration.

  3. The information-propagation and attention diagnostics support the design motivation but do not independently establish a causal explanation. Structural reachability in Figure 3 is computed from the attention-stage input and explicitly excludes the preceding convolutional receptive fields, so it describes only the attention paths as defined. Low attention entropy in Figures 4–5 indicates concentrated weights, but does not directly show that the weights concentrate on spatial neighbors or establish that long-range interactions can be removed.

  4. Some small numerical differences and secondary design conclusions require stronger controls. Neither the 0.2–0.5-point accuracy gains over LitePT-S in the main tables nor the schedule rankings in Table 5 are supported by results across multiple random seeds. Repeating the key controlled comparison in Table 4 should take priority over rerunning every configuration. Table 6 simultaneously changes RoPE axis allocation, attention-head dimension, and parameter count. It therefore supports the accuracy–cost performance of a particular configuration, but does not isolate the effect of axis allocation. A comparison at a fixed head dimension would be more convincing.

References

[1] Yuanwen Yue, Damien Robert, Jianyuan Wang, et al. LitePT: Lighter Yet Stronger Point Transformer. CVPR, 2026.

[2] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, et al. Point Transformer V3: Simpler, Faster, Stronger. CVPR, 2024.

[3] Kadir Yilmaz, Adrian Kruse, Tristan Höfer, Daan de Geus, and Bastian Leibe. Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding. ECCV, 2026.

[4] Jiaxu Wan, Hong Zhang, Ziqi He, et al. SP²T: Sparse Proxy Attention for Dual-stream Point Transformer. ICCV, 2025.

[5] Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D. Barfoot, and Jian Zhang. KPConvX: Modernizing Kernel Point Convolution with Kernel Attention. CVPR, 2024.

Questions for authors
  1. Could you provide an incremental comparison from LitePT-S through attention-stage consolidation, normalization changes and layout reuse, and the introduction of global modules, reporting accuracy, parameter counts, and component timings at each step? If some changes must be applied jointly, please clarify their dependencies.
  2. What are the scene-level latencies of the models under the complete no-TTA fragment-aggregation protocol in Table 2? Please distinguish preprocessing, model forwards, and prediction aggregation; alternatively, provide single-fragment accuracy corresponding to the single-forward timing.
  3. Are the fixed-window versus interleaved-attention comparison in Table 4 and the schedule comparisons with the same number of global modules in Table 5 stable across random seeds? Could you provide distance-conditioned attention statistics or interventions on cross-window interactions to strengthen the diagnostic conclusions?
6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Thank you for your detailed comments. We have carefully considered your feedback and provide additional experimental results below.

Regarding the contributions of individual changes. We have done an incremental ablation from LitePT-S to MixPT-S using three random seeds across all configurations (Table R2-1). Merging the attention stages requires removing the intermediate pooling and the corresponding decoder stage, so we evaluated these changes together.

Table R2-1: Incremental comparison (w/o TTA, mIoU mean ± stdev.)

StepChange from previous stepScanNet #ParamsScanNet mIoU (%)ScanNet Latency (ms)Sem.KITTI #ParamsSem.KITTI mIoU (%)Sem.KITTI Latency (ms)
0LitePT-S: two attention stages: 6 window blocks (width 252) → spatial pooling → 2 window blocks (width 504).12.7M75.2 ± 0.2925.712.7M69.7 ± 0.1736.4
1Merge attention stages: 8 window blocks at the first stage's resolution and width (252); remove the corresponding decoder level.7.8M74.6 ± 0.1921.17.8M69.3 ± 0.0631.0
2Widen the merged stage: width 252→352, heads 14→16, and RoPE (6,6,6)→(8,8,6).13.7M74.9 ± 0.2921.613.6M69.7 ± 0.1233.5
3Remove block-entry LayerNorm; retain attention/MLP pre-norm.13.6M74.9 ± 0.3421.313.6M69.2 ± 0.6833.0
4Fix the serialization order: four shuffled orders → one fixed order (TZ), with layout reuse across all eight blocks.13.6M73.9 ± 0.1517.213.6M68.5 ± 0.0927.9
5Replace two window blocks with global blocks: WWGWWGWW (MixPT-S).13.6M75.9 ± 0.3117.514.1M69.9 ± 0.1428.2
AuthorsOpenReview ↗

Response to Weakness 2 and Question 2

Regarding additional baselines. Following your suggestion, we included SP²T and KPConvX-L as additional baselines (Table R2-2 and Table R2-3, respectively). All latency measurements refer to a single model forward pass on an RTX A6000. MixPT-S shows competitive accuracy with much lower latency than these baselines.

Table R2-2: Semantic segmentation on validation sets (w/ TTA)

MethodSem.KITTI mIoU (%)Sem.KITTI Latency (ms)Waymo mIoU (%)Waymo Latency (ms)nuScenes mIoU (%)nuScenes Latency (ms)
SP²T [4]71.7122.971.9152.881.2107.2
MixPT-S71.228.272.942.881.819.2

Table R2-3: Semantic segmentation on the validation set (w/ TTA)

MethodScanNet mIoU (%)ScanNet Latency (ms)
KPConvX-L [5]76.3150.7
MixPT-S76.917.5

Regarding single-fragment evaluation. We additionally report single-fragment accuracy without TTA in Table R2-4. To clarify, each fragment represents the entire voxelized scene with one representative point per occupied voxel. It does not refer to a small part of the scene. The standard evaluation aggregates predictions across multiple fragments with different representative points to obtain point-level predictions for all original points. For single-fragment evaluation, we use the first point in each voxel and perform a single forward pass per scene. The resulting prediction is assigned to all original points within the same voxel without fragment aggregation. As shown in Table R2-4, accuracy remains comparable to standard evaluation.

Table R2-4: Single-fragment and fragment-aggregated accuracy (w/o TTA)

MethodEvaluationSem.KITTI mIoU (%)Sem.KITTI mAcc (%)Waymo mIoU (%)Waymo mAcc (%)nuScenes mIoU (%)nuScenes mAcc (%)ScanNet mIoU (%)ScanNet mAcc (%)
LitePT-SSingle fragment69.676.771.783.280.487.275.283.3
LitePT-SFragment aggregation69.876.871.782.880.387.175.583.4
MixPT-SSingle fragment69.977.272.083.080.687.275.784.1
MixPT-SFragment aggregation70.077.372.183.080.687.476.084.3
AuthorsOpenReview ↗

Response to Weakness 3 and Question 3 (attention diagnostics)

Regarding reachability and preceding convolutions. Our original analysis in Figure 3 focused on structural reachability within the attention stage, excluding the preceding convolutions. We extended this analysis to include the preceding convolutional stages and pooling operations. In the original analysis, the reachable-token ratios after the first six attention blocks of LitePT-S were 77.6% with four changing serialization patterns and 4.8% with a fixed pattern on the Waymo validation set. Including the preceding operations increased these ratios to 78.8% and 5.6%, respectively. Thus, information propagation under fixed windows remains limited even when preceding convolutions are included.

Regarding spatial locality and long-range interactions. To examine spatial locality, we analyze attention weights by query-key spatial distance on the Waymo validation set (Table R2-5). We report the fraction of keys and attention mass within 5 m. We also report d50d_{50} and d90d_{90}, the median distances across queries enclosing 50% and 90% of attention mass, respectively. Larger values indicate a broader spatial distribution of attention. In VoLT-S, approximately half of the attention mass is concentrated within 5 m at the third block. Later blocks also show spatially concentrated attention. These observations suggest that full global attention may not be necessary at every block.

We further restrict the global attention range of MixPT-S at evaluation time without retraining (Table R2-6). The resulting accuracy degradation supports the contribution of long-range interactions to the trained model's predictions.

Table R2-5: Global attention locality on the Waymo validation set

For blocks 4–12, each range denotes the minimum and maximum of the corresponding metric across these blocks.

MethodBlock indexKeys within 5 m (%)Attention mass within 5 m (%)d50d_{50} (m)d90d_{90} (m)
VoLT-S12.12.234.964.9
VoLT-S22.135.712.040.7
VoLT-S32.150.54.832.8
VoLT-S4–122.136.6–65.32.1–9.316.6–42.5
MixPT-S3 (G)1.13.840.279.5
MixPT-S6 (G)1.110.927.373.7

Table R2-6: Global-attention interventions for MixPT-S on the Waymo validation set (w/o TTA)

Attention radius (m)Full4020105
mIoU (%)72.171.671.169.565.6
Reviewer_81dgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MixPT combines fixed-order serialized window attention with periodically interleaved global attention. For large token sets, global blocks operate on pooled features and return residual updates to the finer stream. Built on LitePT, the architecture consolidates its attention hierarchy and reduces token-reorganization overhead. Validation experiments cover four semantic segmentation benchmarks and two indoor instance segmentation datasets. MixPT-S achieves competitive accuracy with 19–32% lower single-forward latency than LitePT-S, supported by architectural ablations. Larger variants are evaluated, although their additional accuracy gains are modest. No independent test-set results are reported.

Critical strengths
  • Consistent efficiency improvements. MixPT-S reduces forward latency across four datasets while maintaining competitive accuracy. This represents a useful practical contribution.

  • Controlled ablations support adding global attention. Table 4 shows gains of 1.4 mIoU on SemanticKITTI and 2.3 on ScanNet over fixed-window attention, with little additional latency. These comparisons provide substantive support for the core design.

Critical weaknesses
  1. The evidence for beneficial model scaling is weak. Increasing parameters from 14.1M to 89.4M—approximately 6.34×—improves mIoU by only 0.3–0.9 points, while latency increases approximately 1.9–2.8× (Tables 1–2). Waymo shows no improvement from S to B, and TTA results include nonmonotonic gains. These results demonstrate that larger variants can be trained, but offer limited evidence that additional capacity is effectively used. Without scale-specific convergence analysis, controlled depth/width comparisons, or run variability, architectural bottlenecks cannot be distinguished from optimization or data limitations.

  2. Independent evaluation and baseline comparability are unresolved. The results are confined to validation sets, without test-set confirmation after architectural selection; PTv3 reports both validation and test results on several corresponding benchmarks. Moreover, the draft’s PTv3 entries are not sufficiently reconciled with published results. Appendix B.1 mixes original and reproduced baseline scores without identifying each entry’s source. Differences may reflect training, EMA, preprocessing, or TTA rather than incorrect results, but their unresolved provenance weakens confidence in the comparative claims.

Questions for authors
  1. Were B/L models separately tuned and trained to convergence? What learning curves and run variability explain their limited gains, and can controlled depth/width comparisons identify the cause?
  2. Can the selected configurations be evaluated on the available test sets of ScanNet, nuScenes, and SemanticKITTI to confirm the validation conclusions?
  3. For each PTv3 entry, which checkpoint, training-data regime, and evaluation configuration were used? Please identify published versus reproduced results and compare the same checkpoint under both w/o TTA and TTA protocols.
4 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Questions 1–2

Thank you for your thoughtful feedback. It helped us further investigate model scaling and strengthen our evaluation.

Regarding model scaling. We agree that scaling brings only small accuracy gains under single-dataset training. For larger models, we changed the batch size, learning rate, and number of GPUs while keeping the augmentation and regularization settings the same. To better understand these limited gains, we also explored joint training on SemanticKITTI, Waymo, and nuScenes following Volt. On SemanticKITTI, scaling gains are larger under joint training (Table R1-1), suggesting that the available training data may partly explain the small gains under single-dataset training. Jointly trained MixPT-B also shows competitive performance against PTv3+PPT and Volt-B (Table R1-2).

Regarding test-set evaluation. We also evaluated jointly trained MixPT-B on the SemanticKITTI and nuScenes test sets. The performance advantage seen on the validation sets also holds on both test sets (Table R1-2).

Table R1-1: Model scaling on the validation set (w/o TTA, mIoU, %)

Method#ParamsSem.KITTI (Single)Sem.KITTI (Joint)
MixPT-S14.1M70.070.8
MixPT-B51.8M70.471.6
MixPT-L89.4M70.671.9

Table R1-2: Joint-training validation and test results (w/ TTA, mIoU, %)

TrainingMethod#ParamsSem.KITTI (Val)Sem.KITTI (Test)nuScenes (Val)nuScenes (Test)Waymo (Val)
JointPTv3 + PPT46.3M72.375.581.283.072.1
JointVolt-B87.7M72.575.282.282.273.4
JointMixPT-B51.8M72.676.282.383.373.8
AuthorsOpenReview ↗

Response to Weakness 2 and Question 3

Thank you for your comments on the baseline comparisons. We have also tried to make the comparisons fair across methods. For each dataset, we followed the original PTv3 training recipe and applied EMA. All methods were evaluated using the same evaluation protocol. For each method, the same checkpoint was used for evaluation with and without TTA. Table R1-3 distinguishes published and reproduced results, with bold values indicating those used in the current manuscript. Our reproduced TTA results exceeded the published values across all four datasets. The manuscript already reports these results for SemanticKITTI, nuScenes, and ScanNet. We will also update the Waymo result with our reproduced result.

Table R1-3: PTv3 semantic segmentation on validation sets (mIoU / mAcc, %)

PTv3Sem.KITTI (mIoU)Sem.KITTI (mAcc)Waymo (mIoU)Waymo (mAcc)nuScenes (mIoU)nuScenes (mAcc)ScanNet (mIoU)ScanNet (mAcc)
Official (w/ TTA)70.8-71.380.580.4-77.5-
Reproduced (w/ TTA)71.077.971.882.181.287.577.685.9
Reproduced (w/o TTA)69.677.270.781.379.686.676.585.3

Update: We have now replaced the PTv3 TTA result on Waymo with our reproduced result in the revised manuscript (Table A4, p. 19).

Dear Reviewer,

Thank you again for your valuable feedback. To address your concerns about model scaling, we explored joint training across multiple datasets and observed larger scaling gains on SemanticKITTI. This suggests that limited training data may partly explain the small scaling gains under single-dataset training. We also provided our test set results and clarifications on the PTv3 baseline evaluation. These updates are included in the revised manuscript (Tables A4–A6, pp. 19–20).

We would be happy to answer any remaining questions during the discussion period. Thank you for your time.

AuthorsOpenReview ↗

Additional Response to Weakness 1 and Question 1

We also evaluated model scaling by increasing depth and width. As shown in Table R1-4, these changes resulted in small accuracy gains on SemanticKITTI under single-dataset training. Together with the larger gains under joint training (Table R1-1), these results suggest that the limited amount of training data may partly explain the small improvements.

Table R1-4: Depth and width scaling on SemanticKITTI (w/o TTA)

Method# Encoder blocks (0–3)Encoder channels (0–3)#ParamsmIoU (%)Latency (ms)
MixPT-S(2, 2, 2, 8)(36, 72, 144, 352)14.1M70.028.2
+ Depth(3, 3, 3, 14)(36, 72, 144, 352)24.3M70.439.7
+ Width(2, 2, 2, 8)(54, 108, 216, 512)30.1M70.136.2
MixPT-B(3, 3, 3, 14)(54, 108, 216, 512)51.8M70.456.4
+ Depth(3, 3, 3, 24)(54, 108, 216, 512)85.4M70.775.7
+ Width (MixPT-L)(3, 3, 3, 14)(72, 144, 288, 672)89.4M70.672.4

Comments on the paper

3 comments

First of all, we really appreciate your time and effort in serving as our Area Chair. We also thank all reviewers for their careful reading and constructive feedback. All comments have helped us clarify our contributions and improve our experimental analysis. We summarize our responses below:

Reviewer 81dg (R1): Model scaling, test-set evaluation, and PTv3 baseline reproduction.

Reviewer 26z9 (R2): Architectural ablations, baseline and evaluation comparisons, multi-seed robustness, RoPE allocation, and attention diagnostics.

Reviewer AVNP (R3): Conceptual novelty, architectural contributions, and accuracy-latency trade-offs.

Tables are numbered separately for each reviewer (e.g., Table R1-1 refers to the first table for Reviewer 81dg). Table numbers without an R-prefix in the reviews refer to the original manuscript.

AuthorsOpenReview ↗

Summary of Manuscript Revisions

We sincerely thank the reviewers for their helpful comments. We have updated the manuscript to include additional experiments and analyses. The updated content is highlighted in blue and the relevant reviewers are listed in the right margin of the revised PDF.

Changes include:

  • Multi-seed ablations (Tables 4–5, p. 8) [Reviewer 26z9]
  • PTv3 evaluation (Table A4, p. 19) [Reviewer 81dg]
  • Joint training and test-set evaluation (Tables A5–A6 and discussion, p. 20) [Reviewer 81dg]
  • Single-fragment evaluation (Table A8 and discussion, p. 20) [Reviewer 26z9]
  • Incremental comparison (Table A9 and discussion, p. 21) [Reviewers 26z9, AVNP]
  • Global attention analysis (Figure A2 and discussion, p. 22) [Reviewer 26z9]
  • Minor clarifications: captions (Figure 1, p. 1; Table A4, p. 19) and labels (Figures A3–A5, pp. 23–25).
AuthorsOpenReview ↗

Summary of rebuttal

Dear Area Chair and Reviewers,

We sincerely thank all of you for your time and consideration. Here are the main points discussed during the discussion period.

The strengths of our work highlighted by the reviewers:

  • Consistent efficiency improvements with competitive accuracy. MixPT-S achieves 19–32% lower latency than LitePT-S (Table 1, p. 6) while maintaining competitive accuracy compared with representative point transformer based baselines (Table 2, p. 6). [Reviewers 81dg, 26z9, AVNP]
  • Experimental support for the key design choices. Ablations show that using a fixed serialization layout reduces latency while periodic global attention improves accuracy with only a small increase in latency. [Reviewers 81dg, 26z9, AVNP]
  • Detailed implementation and evaluation settings. [Reviewer 26z9]

Our responses to the reviewers:

Contributions and multi-seed robustness.
We provided a step-by-step comparison from LitePT-S to MixPT-S (Tables R2-1 and R3-1) and repeated the requested ablations over three random seeds (Tables R2-7–R2-8). [Reviewers 26z9, AVNP]

Model scaling and test set results.
We further examined depth and width scaling (Table R1-4) and found larger scaling gains on SemanticKITTI under joint training (Table R1-1). Jointly trained MixPT-B outperformed the other jointly trained baselines on the SemanticKITTI and nuScenes validation sets, and this also held on both test sets (Table R1-2). [Reviewer 81dg]

Baseline evaluation.
We clarified the PTv3 evaluation (Table R1-3), added baseline comparisons (SP²T: Table R2-2; KPConvX-L: Table R2-3) and single-fragment accuracy results (Table R2-4). [Reviewers 81dg, 26z9]

Attention and RoPE analyses.
We extended our analyses of attention reachability, global attention (Tables R2-5–R2-6), and RoPE axis allocation (Table R2-9). [Reviewer 26z9]

The revisions in the manuscript are highlighted in blue, and the relevant reviewers are listed in the right margin.

Thank you all again for your valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.