acceptodds
Under review as a conference paper at ICLR 2027

GSJEPA: Self-Supervised Approach to Learning Gaussian Splat Representations Using Joint Embedding Predictive Architecture

Abstract

We introduce a self-supervised representation learning framework for 3D objects represented as Gaussian Splats based on Joint Embedding Predictive Architectures (JEPA). GS-JEPA represents individual Gaussian primitives as nodes in a local graph and learns their representations through latent prediction, explicitly modeling geometric and appearance-aware relationships between primitives. To isolate the effect of relational modeling, we additionally consider a point-based formulation that treats Gaussian primitives as an unordered set and serves as a baseline. Both formulations learn representations directly from Gaussian splat objects without relying on reconstruction objectives. Across classification and segmentation benchmarks, predictive learning outperforms the reconstruction-based Gaussian Masked Autoencoder. Notably, despite being pretrained exclusively on synthetic objects, the proposed JEPA-based approaches generalize effectively to real reconstructed Gaussian splat data. On MACGS classification, the point-based and graph-based approaches achieve and accuracy, respectively. On ShapeNetPart segmentation, the graph-based model achieves an mIoU of , demonstrating the potential benefits of explicit relational modeling. These results establish JEPA-based predictive learning as an effective approach for transferable Gaussian splat representations and highlight relational modeling as a promising direction for 3D Gaussian representation learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-supervised Learning, Gaussian Splatting, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ioooOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents GS-JEPA, a self-supervised pretraining method for Gaussian Splatting that replaces masked reconstruction of Gaussian attributes with JEPA-style latent prediction. Two variants of the method are presented.

(i) Point-GS-JEPA, where gaussian primitives are grouped into patches, and then passed to a PointNet-based tokenizer to produce patch embeddings. Then, a JEPA-style training is done by predicting representations of masked patches.

(ii) Graph-GS-JEPA, where each Gaussian is treated as a node in a KNN graph with geometric and appearance edge features and uses a Graphormer-style encoder.

The models are pretrained on synthetic objects from ShapeSplat and evaluated on the tasks of object classification and part segmentation.

Critical strengths
  • The paper is clear, well-written, and easy to follow. The progression from a simpler baseline (Point-GS-JEPA) to the full method (Graph-GS-JEPA) makes the logical development of the method particularly clear.
  • The paper presents an interesting application of JEPA-style objectives to Gaussian representations.
  • The experiments are comprehensive and well structured.
  • The results showcase the usefulness of the JEPA-style pretraining, particularly under distrubution shifts.
Critical weaknesses
  • In table 2, Point-GS-JEPA beats Graph-GS-JEPA on MACGS data under all three protocols (linear evlauation, MLP evaluation, and full fine-tuning). This means that the graph-based method is less transferable than the point-based one.
  • The ablation in Table 5 shows limited gains under full fine-tuning. Under linear probing on MACGS the configuration without any edge attributes is best.
  • The segmentation advantage of Graph-GS-JEPA cannot be directly attributed to relational modeling, since the two variants also differ in the tokenization procedure, backbone, predictor and parameter count. Without this evidence, the contribution largely reduces to applying Point-JEPA to Gaussian attributes, and the abstract's claims about relational modeling can be seen as overstated.
  • My points above lead to the fact that the benefit of graph-based relational modeling is not established.
Questions for authors
  • Can the authors provide more intuition in regards to why Graph-GS-JEPA transfer worse than Point-GS-JEPA on MACGS?
  • How can one explain the decrease in performance when adding edge attributes under linear probing on MACGS?
  • Could the authors report the edge-attribute ablation for segmentation with the Graphormer backbone, as done in Figure 8 for GINE?
  • Could the authors evaluate Point-GS-JEPA with smaller patches, to test whether the segmentation gains come from finer granularity rather than relational modeling?
3 comments
AuthorsOpenReview ↗

Addressing review (1/2)

Thank you for your careful analysis of transferability, edge attributes, and the segmentation comparison.

W1. "In table 2, Point-GS-JEPA beats Graph-GS-JEPA on MACGS data under all three protocols (linear evlauation, MLP evaluation, and full fine-tuning). This means that the graph-based method is less transferable than the point-based one."

Both Point-GS-JEPA and Graph-GS-JEPA (Graphormer) outperform Gaussian-MAE on both classification datasets across all three evaluation protocols. With frozen backbones, MACGS linear-probing accuracy improves from 81.55% to 90.52% and 86.67%, respectively. We acknowledge that Point-GS-JEPA performs better on MACGS and position the graph formulation as a parameter-efficient alternative, rather than a uniformly superior replacement. It uses approximately 45% fewer trainable pretraining parameters and achieves stronger segmentation results. We will make this tradeoff explicit in the revised manuscript.

W2. "The ablation in Table 5 shows limited gains under full fine-tuning. Under linear probing on MACGS the configuration without any edge attributes is best."

Table 5 suggests that the benefits of edge attributes depend on the transfer setting and the characteristics of the downstream dataset. Under frozen linear probing, edge-dependent features learned on synthetic ShapeSplat may transfer less effectively to real-world MACGS, explaining why the configuration without edge attributes achieves the best performance.

We hypothesize that this reflects differences in local geometric consistency. Gaussian reconstruction is primarily optimized for rendering quality, so Gaussians representing the same local object structure need not be spatially close. Such relationships may be less consistent in MACGS than in the synthetic pretraining data, making learned edge attributes less reliable under transfer.

Full fine-tuning allows the model to adapt these representations to the target distribution, resulting in modest improvements from edge attributes. In contrast, on ModelSplat40, which is closer to the pretraining distribution, using all edge attributes achieves the best linear-probing performance. We will include this discussion in the revised manuscript, presenting the explanation as a hypothesis and clarifying that removing edge attributes still retains the KNN topology.

W3. "The segmentation advantage of Graph-GS-JEPA cannot be directly attributed to relational modeling, since the two variants also differ in the tokenization procedure, backbone, predictor and parameter count. (...) the abstract's claims about relational modeling can be seen as overstated."

Point-GS-JEPA serves as an architecture-matched baseline for comparison with Gaussian-MAE, while the graph formulation focuses on local relationships between individual Gaussians. Our initial approach used Point-GS-JEPA with one Gaussian per patch. Although this preserved primitive-level granularity, it was computationally expensive and led to representation collapse during pretraining. This motivated Graph-GS-JEPA, which combines fine-grained Gaussian representations with local graph interactions and a cross-attention predictor to reduce predictor attention complexity while preserving primitive-level information.

Our intuition is that Gaussians belonging to different semantic parts may exhibit differences in geometry and appearance. Encoding pairwise attribute differences along graph edges could help identify potential part boundaries and distinguish interactions within a part from those across parts. These cues may benefit segmentation. However, the two formulations also differ in tokenization, backbone, predictor, and parameter count. We will revise the Abstract and segmentation discussion to distinguish this architectural motivation from a causal attribution of the gains solely to relational modeling.

W4. "My points above lead to the fact that the benefit of graph-based relational modeling is not established."

Our results support a compact graph formulation with competitive classification and stronger segmentation among the evaluated Gaussian-based methods. The architecture-matched Point-GS-JEPA/Gaussian-MAE comparison provides clearer evidence for predictive pretraining. We will distinguish these findings from the unresolved contribution of individual graph components.

AuthorsOpenReview ↗

Addressing review (2/2)

Q1. "Can the authors provide more intuition in regards to why Graph-GS-JEPA transfer worse than Point-GS-JEPA on MACGS?"

Graph-GS-JEPA explicitly models local Gaussian neighborhoods, which may make its representations more sensitive to changes in Gaussian density, geometry, and reconstruction artifacts between synthetic ShapeSplat and real-world MACGS. Point-GS-JEPA's patch aggregation and global attention may be less sensitive to these local variations. This interpretation is consistent with Table 5: edge attributes are less beneficial on MACGS when the encoder is frozen, while fine-tuning allows their contribution to adapt. Nevertheless, Graph-GS-JEPA outperforms Gaussian-MAE on MACGS with substantially fewer trainable parameters, supporting its value as a compact alternative while highlighting robustness to distribution shift as an opportunity for improvement. We will present this interpretation as a hypothesis.

Q2. "How can one explain the decrease in performance when adding edge attributes under linear probing on MACGS?"

As discussed in W2, edge-dependent features learned on synthetic ShapeSplat may be less reliable on real-world MACGS because local Gaussian relationships differ between the datasets. A frozen encoder cannot adapt these features to the target distribution, whereas full fine-tuning permits adaptation and yields modest improvements. This suggests sensitivity to distribution shift, rather than a consistent benefit from adding edge attributes.

Q3. "Could the authors report the edge-attribute ablation for segmentation with the Graphormer backbone, as done in Figure 8 for GINE?"

The Graphormer segmentation ablation shows that adding distance, direction, and opacity relationships improves both category and instance mIoU. The table below reports the values from the chart; attributes are added cumulatively.

Edge attributesCategory mIoU (%)Instance mIoU (%)
None84.1585.59
Distance84.6585.79
+ Direction84.8785.98
+ Opacity (Vis)85.1586.27
+ Scale85.0086.11
+ Rotation84.9786.00
+ Appearance84.9986.03

Distance, direction, and opacity together improve category/instance mIoU by 1.00/0.68 percentage points over the configuration without edge attributes. Further attributes do not improve the result, suggesting that selected local relationships are more useful than including every descriptor. Because the backbone is held fixed, this comparison directly examines the contribution of edge attributes within Graphormer. We will include this table and the updated discussion in the revised manuscript.

Q4. "Could the authors evaluate Point-GS-JEPA with smaller patches, to test whether the segmentation gains come from finer granularity rather than relational modeling?"

We explored the finest-granularity setting, assigning one Gaussian to each Point-GS-JEPA token, but observed representation collapse during pretraining. Thus, increasing granularity alone did not yield a viable alternative under the tested setup. Graph-GS-JEPA retains Gaussian-level tokens while providing explicit neighborhood interactions. This motivates its design, although the collapse at single-Gaussian granularity does not rule out benefits from intermediate patch sizes or establish relational modeling as the sole source of segmentation gains.

Additionally, as requested, we conducted further experiments to investigate the influence of group size on segmentation and MACGS classification using linear probing, while keeping the number of groups fixed.

For MACGS classification, we observe a consistent decrease in accuracy as the group size decreases, potentially due to smaller local receptive fields capturing less geometric context and being more sensitive to noise.

In contrast, the effect on segmentation performance is less straightforward. Reducing the group size from 32 to 16 yields a slight improvement in both mIoU_I and mIoU_C, potentially because smaller local neighborhoods preserve finer-grained geometric details that benefit segmentation. However, further reducing the group size to 8 does not improve performance, suggesting that excessively small neighborhoods may provide insufficient local context. Having in mind all those observations and results we do not believe that high segmentation results on Graph GS-JEPA comes solely from the granular representations, but rather whole architecture and modeling approaches as the whole.

Group sizemIoU_ImIoU_CMACGS-LINEAR-ACC
883.6085.4488.06
1683.9785.7189.98
3283.8585.5690.52

I thank the authors for their detailed response and the additional experiments.

  • The ablation and the smaller patch Point-GS-JEPA study are helpful. I also appreciate that the authors will distinguish the evidence for JEPA-style predictive pretraining from the less certain contribution of graph-based relational modeling.
  • My mains concerns are reduced but not fully resolved.
    • Point-GS-JEPA and Graph-GS-JEPA still differ in tokenization, backbone, and parameter count. As such, the current experiments do not isolate the benefit of relational modeling.
    • The provided explanation for the weaker transfer is a plausible hypothesis rather than a controlled finding.

Overall, the rebuttal improves the paper and appropriately narrows its claims, but I would maintain my current assessment.

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

They take Point-JEPA and run it on 3DGS objects. Point-GS-JEPA = treat gaussians as points w/ extra attrs (FPS/KNN, PointNet tokenizer, transformer student + EMA teacher, smooth L1). Graph-GS-JEPA puts a KNN graph on primitives, Graphormer-ish encoder, handmade edges (dist / dir / opacity / scale / rot / SH), sample connected target subgraphs, 6-block cross-attn predictor. Pretrain on synthetic ShapeSplat, then ModelSplat40 + reconstructed MACGS cls (linear / mlp3 / full ft, 3 seeds) and ShapeNetPart via pushing gaussian feats back to points

Critical strengths
  • setup is cleaner than a lot of “SSL but on the new 3d toy” papers. Point-GS-JEPA shares tokenizer/encoder with Gaussian-MAE so the MACGS full-ft gap (95.76 vs 93.31) is at least about the objective. they also have from-scratch Graphormer / Point Transformer. and they actually drop target nodes+edges so the graph cant leak, good
  • they dont hide the mixed results. graph helps on ModelSplat40 and part seg, the simpler point version is better on noisy MACGS. they say its object-level only
Critical weaknesses
  • kinda just “does Point-JEPA still work if the points have gaussian attrs”. the graph model is the only new bit and it loses on the more interesting shift (synth -> real reconstructed). so the title/abs pushing “relational modeling” feels off
  • gains are small and baselines are thin. ModelSplat40 full ft +0.51 over Gaussian-MAE. ShapeNetPart 85.15/86.27 vs MAE and vs old Point-M2AE (84.9/86.5) not a big story. they cite GS-PT / GaussianPretrain but never run them. no scene task. “new evaluation benchmark” is just linear/ft on existing ShapeSplat/MACGS
Questions for authors
  1. why does the graph version lose on MACGS. topology vs noise, or just capacity. injecting noise on ShapeSplat would help, right now this is post-hoc
  2. please put GS-PT or GaussianPretrain (or any non-MAE gaussian ssl) on the same MACGS proto. otherwise you only showed JEPA > this one MAE
  3. how do you normalize the gaussian attrs in the PointNet tokenizer. fixed SH degree? drop view-dep SH or not
  4. ShapeNetPart — how do you assign gaussian feats back to labeled points, 1-1 subsample or interpolate
  5. Table 1 graph model is 13.5M vs 24.7M. is MACGS just an under-capacity compare. width-matched run?
2 comments
AuthorsOpenReview ↗

Addressing review (1/2)

Thank you for your detailed feedback on our contribution, comparisons, and evaluation protocol.

W1. "kinda just ‘does Point-JEPA still work if the points have gaussian attrs’. the graph model is the only new bit and it loses on the more interesting shift (synth -> real reconstructed). so the title/abs pushing ‘relational modeling’ feels off"

Point-GS-JEPA intentionally serves as an architecture-matched baseline for comparison with Gaussian-MAE. The graph architecture was motivated by experiments with single-Gaussian Point-GS-JEPA patches, which were computationally expensive and led to representation collapse. Graph-GS-JEPA combines fine-grained Gaussian representations with local graph interactions and fewer trainable parameters. Although it performs below Point-GS-JEPA on MACGS, it outperforms Gaussian-MAE across the main classification protocols and achieves stronger segmentation results than both methods. We will revise the Abstract to reflect these tradeoffs without attributing the gains solely to relational modeling.

W2. "gains are small and baselines are thin. ModelSplat40 full ft +0.51 over Gaussian-MAE. ShapeNetPart 85.15/86.27 vs MAE and vs old Point-M2AE (84.9/86.5) not a big story. they cite GS-PT / GaussianPretrain but never run them. no scene task. ‘new evaluation benchmark’ is just linear/ft on existing ShapeSplat/MACGS"

The clearest improvements emerge on MACGS: linear-probing accuracy increases from 81.55% for Gaussian-MAE to 90.52% and 86.67% for the two JEPA formulations. The ModelSplat40 fine-tuning gain is modest. On ShapeNetPart, Graph-GS-JEPA improves category/instance mIoU over Gaussian-MAE by 0.92/0.39 percentage points, but does not exceed Point-M2AE on instance mIoU.

GS-PT and GaussianPretrain address different input and supervision settings (see Q2). By “new evaluation benchmark,” we intended an extension of Gaussian-MAE's synthetic-data evaluation to synthetic-to-real transfer, rather than a new dataset. We will clarify this wording and explicitly state that our experiments cover object-level tasks, without scene-level evaluation. Segmentation additionally demonstrates transfer from Gaussian-derived features to labeled ShapeNetPart points despite their different primitive distributions.

Q1. "why does the graph version lose on MACGS. topology vs noise, or just capacity. injecting noise on ShapeSplat would help, right now this is post-hoc"

A plausible explanation is that spatial proximity is a less reliable indicator of semantic similarity in real Gaussian reconstructions. Gaussians are optimized for rendering fidelity, and incomplete observations and reconstruction ambiguities may change local neighborhoods. Graph-GS-JEPA explicitly propagates information along local edges, potentially making it more sensitive to these changes than Point-GS-JEPA's patch aggregation and global attention. This interpretation is consistent with the edge-attribute ablation in Table 5. Simple noise injection would not necessarily reproduce these structured reconstruction effects; we will distinguish this hypothesis from the empirical findings in the revised manuscript.

Q2. "please put GS-PT or GaussianPretrain (or any non-MAE gaussian ssl) on the same MACGS proto. otherwise you only showed JEPA > this one MAE"

GS-PT uses Gaussian splatting to support multimodal pretraining of point-cloud encoders, while GaussianPretrain learns visual representations from multi-view images with LiDAR-guided Gaussian reconstruction. Our method directly encodes existing Gaussian primitives and their attributes, making Gaussian-MAE the closest direct comparison. We will clarify these differences in input and supervision and frame our empirical conclusion as an improvement over Gaussian-MAE under matched evaluation protocols.

AuthorsOpenReview ↗

Addressing review (2/2)

Q3. "how do you normalize the gaussian attrs in the PointNet tokenizer. fixed SH degree? drop view-dep SH or not"

Normalization follows Gaussian-MAE. Positions are centered and divided by the maximum distance from the centroid. Gaussian scales are converted from logarithmic to linear values and divided by the same spatial factor, then mean-centered and divided by their maximum Euclidean norm. Opacities are passed through a sigmoid and mapped to [-1, 1]. Quaternions are normalized to unit length, with a nonnegative first component. We retain only the three degree-0 SH coefficients and discard higher-order, view-dependent coefficients. We will add these details to the revised manuscript.

Q4. "ShapeNetPart — how do you assign gaussian feats back to labeled points, 1-1 subsample or interpolate"

We use spatial interpolation, not a one-to-one correspondence. Gaussians and labeled points are sampled independently, aligned in the same coordinate system, and normalized using the same centroid and spatial scale. Features are interpolated from the three nearest Gaussian feature locations (group centers for patch features), using normalized inverse-squared-distance weights. The interpolated feature is combined with point coordinates and processed by the segmentation head. We will add this procedure to Section 4.2.

Q5. "Table 1 graph model is 13.5M vs 24.7M. is MACGS just an under-capacity compare. width-matched run?"

Both encoders use the same hidden dimension of 384; parameter counts differ because of depth and architecture. Point-GS-JEPA uses the same 12-layer Transformer encoder as Gaussian-MAE. Increasing graph depth did not improve performance in our explored configurations; it also expands the receptive field and may dilute local distinctions. Nevertheless, these observations do not replace a parameter-matched comparison, and capacity remains a possible contributor to the MACGS gap.

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

This paper explores self-supervised representation learning for 3D Gaussian Splatting (3DGS) using a Joint-Embedding Predictive Architecture (JEPA). The authors introduce two variants. The first is Point-GS-JEPA, which adapts Point-JEPA to take 14-dimensional Gaussian descriptors as input. The second is Graph-GS-JEPA, which represents each Gaussian primitive as a node in a KNN graph, incorporates multi-attribute edge features, and uses a local graph transformer encoder with a cross-attention predictor. The models are pre-trained on ShapeSplat and evaluated on downstream classification (ModelSplat40 and MACGS) and part segmentation (ShapeNetPart).

Critical strengths
  1. Some interesting findings: Point-GS-JEPA shares the same tokenizer and encoder architecture with Gaussian-MAE. This provides a clean controlled comparison between latent prediction and input reconstruction. In particular, Table 10 in Appendix I shows a interesting result on the real-world dataset. While using only xyz coordinates yields nearly identical performance between MAE and JEPA, adding full Gaussian attributes degrades frozen MAE performance but significantly boosts JEPA. This supports the hypothesis that reconstructing noisy low-level Gaussian attributes harms cross-domain transfer, whereas latent prediction avoids this issue.

  2. Proper handling of graph information leakage and quaternion distance. In Graph-GS-JEPA, the authors remove target nodes and all edges connecting context and target subgraphs during pre-training. This effectively prevents message-passing leakage in graph self-supervised learning. The formulation of quaternion angular distance also properly accounts for the double-cover property of rotations.

Critical weaknesses
  1. Weak empirical support for the graph method. The paper lacks a Graph-MAE baseline to isolate the benefit of graph modeling from the JEPA objective. More importantly, Graph-GS-JEPA underperforms the simple Point-GS-JEPA baseline on the real-world MACGS benchmark in both linear probing and fine-tuning. The improvement on synthetic ModelSplat40 is small and within random noise.

  2. Missing sampling and augmentation details. Table 4 shows that objects contain 20k to 110k Gaussians, but Appendix B indicates that only 1024 points are used. The subsampling method is omitted, making it unclear whether KNN graphs on such a small subset preserve true local relations. Furthermore, Appendix F mentions random scaling and rotation of coordinates, but does not state whether scale and quaternion attributes are transformed consistently.

  3. Literature and baseline citation errors. The text repeatedly confuses point2vec (Knaebel et al., 2023) with Points2Vec (Bachmann et al., 2021). Bachmann et al. is incorrectly cited in the Introduction, Section 4.1, and Section 4.2 for masked autoencoding, feature pooling, and segmentation protocols. In Table 2, Point Transformer (Zhao et al., 2021) is cited as an identical baseline, but Zhao uses local vector attention rather than a standard ViT encoder.

Questions for authors

Given that ModelSplat40 and MACGS objects average 22k and 110k Gaussians respectively, how many primitives are fed into Graph-GS-JEPA during pre-training and evaluation, and what sampling strategy is used?

Table 4 shows the same object counts for validation and test sets. Were design choices and hyperparameters tuned directly on the test set?

1 comment

Thank you for your careful review and for highlighting the empirical and methodological points that need clarification.

W1. "The paper lacks a Graph-MAE baseline to isolate the benefit of graph modeling from the JEPA objective. (...) Graph-GS-JEPA underperforms the simple Point-GS-JEPA baseline on the real-world MACGS benchmark (...). The improvement on synthetic ModelSplat40 is small and within random noise."

Both Point-GS-JEPA and Graph-GS-JEPA (Graphormer) outperform Gaussian-MAE on both classification datasets across all three evaluation protocols. Crucially, with the backbone fully frozen, MACGS linear-probing accuracy improves from 81.55% to 90.52% and 86.67%, respectively.

We acknowledge that Point-GS-JEPA performs better on MACGS and position the graph formulation as a parameter-efficient alternative, rather than a uniformly superior replacement. It uses approximately 45% fewer trainable pretraining parameters than Point-GS-JEPA and achieves stronger segmentation results. We will reflect this distinction in the revised manuscript and acknowledge that a Graph-MAE control is needed to isolate the graph architecture from the pretraining objective.

W2. "The subsampling method is omitted, making it unclear whether KNN graphs on such a small subset preserve true local relations. (...) Appendix F mentions random scaling and rotation of coordinates, but does not state whether scale and quaternion attributes are transformed consistently."

Following the Gaussian-MAE evaluation protocol (lines 343–344), we use the same random subsampling strategy for both GS-JEPA approaches, selecting a fixed number of Gaussian primitives for pretraining, 4,096 for ModelSplat40 classification, and 1,024 for segmentation. We will explicitly document this sampling procedure and the input counts for each task. KNN graphs are constructed on the sampled primitives, so they describe the sampled neighborhoods rather than guaranteeing preservation of the original dense object's local relations.

Augmentations jointly update the Gaussian geometry: scaling factors are applied to both positions and scale attributes, while rotations update both positions and orientation quaternions through quaternion composition. Unit-sphere normalization also rescales the Gaussian scales. We will include the updated preprocessing section in the revised manuscript's appendix.

W3. "The text repeatedly confuses point2vec (Knaebel et al., 2023) with Points2Vec (Bachmann et al., 2021). (...) In Table 2, Point Transformer (Zhao et al., 2021) is cited as an identical baseline, but Zhao uses local vector attention rather than a standard ViT encoder."

We inadvertently cited Points2Vec (Bachmann et al., 2021) where point2vec (Knaebel et al., 2023) was intended. We will correct the references in Sections 4.1–4.2 and revise the Introduction to distinguish reconstruction-based methods from latent predictive learning.

Similarly, the Zhao et al. (2021) citation attached to the supervised Transformer baseline was incorrect. The experiment used the intended architecture-matched encoder trained from random initialization, not Zhao et al.'s local vector-attention model. The error therefore concerns the baseline's name and citation rather than the experimental comparison. We will correct both in the revised manuscript.

Q1. "Given that ModelSplat40 and MACGS objects average 22k and 110k Gaussians respectively, how many primitives are fed into Graph-GS-JEPA during pre-training and evaluation, and what sampling strategy is used?"

As detailed in W2, following the Gaussian-MAE evaluation protocol, we use the same random subsampling strategy for both GS-JEPA approaches: a fixed-size input for pretraining, 4,096 primitives for ModelSplat40 classification, and 1,024 for segmentation. Table 4 reports the full objects' Gaussian counts, not the model input sizes. We will explicitly document the pretraining and MACGS input counts alongside the sampling procedure in the revised manuscript.

Q2. "Table 4 shows the same object counts for validation and test sets. Were design choices and hyperparameters tuned directly on the test set?"

We follow the benchmark evaluation protocols used by Gaussian-MAE and related point-cloud SSL methods (Point-MAE, Point-BERT, etc.), reporting final metrics on the test split. Table 4 uses a combined “Test/Val” heading for the benchmark evaluation split, which we will clarify to avoid implying two independent splits. The supplementary experiments are ablations intended to characterize the effects of architectural and hyperparameter choices, rather than establish a test-selected optimal configuration. These ablations were nevertheless evaluated on the test splits; we will clarify which design choices they informed in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.