acceptodds
Under review as a conference paper at ICLR 2027

Emergent Multi-View Geometry Through Self-Distillation

Abstract

Over a century ago, Henri Poincaré argued that a motionless observer cannot acquire the notion of space. Yet, most visual representation learning methods operate on individual images, while those that leverage multiple views rely on RGB reconstruction, entangling geometry with appearance. We propose Poincar3, a self-supervised method that learns representations from multiple views through self-distillation instead of RGB reconstruction. We combine masked patch and image-level distillation with a teacher that observes additional views, enabling training from scratch without explicit 3D supervision. Poincar3 outperforms both previous single and multi-view self-supervised approaches such as DINOv3, MuM, and Muskie on correspondence estimation, camera pose estimation, and 3D reconstruction. Using a lightweight Poincaré adapter, we also find that our learned features encode camera motion more accurately than existing self-supervised representations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, 3D vision, multi-view learning, representation learning, multiple-view geometry

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_B1pyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Poincar3, a self-supervised multi-view masked modeling framework for learning transferable 3D-aware features from an image sequence of a 3D scene. Poincar3 adopts a multi-view iBOT objective that encourages a student network to mimic the features from a teacher network despite patch-wise masking. Moreover, Poincar3 leverages an additional image-level objective and suggests that the teacher network should see more views than the student, which bring performance improvements.

Putting all together, Poincar3 learn strong multi-view features compared to previous single-view (e.g., DINO) and multi-view baselines, on zero-shot cross-view matching, feedforward reconstruction, as a (much) smaller computational cost. The ablation study clearly shows the contribution of each component in the proposed approach.

Critical strengths

(1) The proposed method, Poincar3, is conceptually simple and clean, making it easy for potential follow-ups and further developments.

(2) Empirical results show that Poincar3 learns transferable 3D-aware multi-view features, demonstrating better performance on multi-view matching and feedforward 3D reconstruction than previous single-view (e.g., DINO) and multi-view self-supervised methods (e.g., MuM). Importantly, this is achieved with much smaller computational costs than state-of-the-art models, e.g., DINO.

(3) The ablation study cleanly attributes the contribution of each proposed component.

Critical weaknesses

(1) Insufficient comparison. Poincar3 primarily demonstrates "3D" downstream tasks, such as cross-view matching and feedforward reconstructions, and it achieves inferior performance in terms of semantic understanding (e.g., ImageNet results) compared to single-view backbone. Given the focus of 3D-related tasks, the proposed method should be compared with the family of method that conducts multi-view self-supervised learning with novel-view synthesis (e.g., RayZer, E-RayZer, and their recent follow-ups), which also focus on 3D-related tasks.

(2) The scope of claim and analysis. The title of "Emergent Multi-View Geometry Through Self-Distillation" feels a bit too grand to me. The strongest evidence for the "emergent multi-view geometry" is from Section 4.3, where the difference of the features from different frames are trained to predict pose changes. The results from Tab. 3 show that previous methods, e.g., MuM, also work reasonably well, and even DINO does a reasonable job on this specific task.

First, I'd like to see if the performance on 3D-related task correlates well with this task (i.e., is this a useful task to indicate performance). Concretely, can we have an updated version of Table 4 that adds this R^2 metric? I'd also like to know more (intuitive or non-intuitive) explanations about why this specific task is a good measure for "emergent multi-view geometry".

Second, while empirically results, including comparison on 3D-related task and ablation study, provide evidence for a strong multi-view encoder, the analysis of why this happened is missing. For example, why is iBOT objective more preferred compared to DINOv2 in terms of 3D-related task? Why semantic understanding is much worse compared with DINO? I believe the answer to these questions may make this submission more complete and comprehensive.

(3) In terms of single-view backbone evaluation (e.g., for DINO), I wonder have the authors tried Depth Anything 3 style DINO adaptation, directly rearranging tokens so that DINO can itself reason over multi-view tokens. I think that would also be a reasonable baseline for "what if we directly adopt a single-view backbone into multi-view scenario".

Questions for authors

I primarily expect the authors explanation about "emergent multi-View geometry" and any insights about how the objective of learning affects performance.

4 comments
AuthorsOpenReview ↗

Response to Reviewer B1py

Comment [1/2]: We thank reviewer B1py for the careful reading of our paper and for helping improve it. The rebuttal is split into two comments due to space limitations. We are also happy to address any further concerns of the reviewer during the discussion period.

Insufficient comparisons

Given the focus of 3D-related tasks, the proposed method should be compared with the family of method that conducts multi-view self-supervised learning with novel-view synthesis (e.g., RayZer, E-RayZer, and their recent follow-ups), which also focus on 3D-related tasks.

We agree that these methods provide valuable additional baselines. We therefore added RayZer, E-RayZer, and xFactor (already cited in the manuscript) to the multi-view correspondence (Tab. 2) and feed-forward reconstruction (Tab.1) evaluations.

First, the results of multi-view correspondence estimation using feature nearest-neighbor matching:

MethodScanNet 5pxScanNet 10pxScanNet 25pxScanNet 50pxNAVI 5pxNAVI 10pxNAVI 25pxNAVI 50px
RayZer20.721.123.829.913.214.219.833.7
E-RayZer21.624.130.939.914.317.428.744.7
xFactor22.828.240.952.515.220.134.048.1
Poincar327.946.279.589.619.233.768.784.2

Secondly, feed-forward reconstruction (Tab. 1):

MethodRE10K @3°RE10K @30°ScanNet++ @3°ScanNet++ @30°MegaDepth @3°MegaDepth @30°ETH3D Acc.ETH3D NCDTU Acc.DTU NC
Head only
DINOv30.018.90.09.41.555.51.180.5910.760.54
Muskie1.136.30.019.00.146.91.120.6011.430.55
E-RayZer2.454.00.06.30.08.31.150.5510.850.54
xFactor4.160.90.332.60.015.91.110.5512.780.51
Poincar32.451.90.148.73.767.50.820.6510.600.59
Transformer
E-RayZer0.830.20.012.60.125.71.060.5911.410.53
xFactor1.846.20.125.00.119.11.150.5910.750.52
Poincar35.262.21.352.15.269.50.750.6810.520.61

For these experiments, we tried multiple configurations to see what part of the models (pose encoder or geometry decoder) that features best are extracted from. For the feed-forward reconstruction experiment, we use the camera tokens from the pose encoders and the dense features from the geometry decoders while for the correspondence estimation we simply use the best perfoming of the two. RayZer was excluded from the feed-forward reconstruction comparison due to computational constraints (it was also the weakest of the compared methods in the correspondence evaluation).

Overall, Poincar3 provides substantial improvements on both correspondence estimation and feed-forward reconstruction. xFactor's camera tokens perform particularly well on RE10K pose estimation.

Title of the paper

The title of "Emergent Multi-View Geometry Through Self-Distillation" feels a bit too grand to me.

We agree with the reviewer that an alternative title might fit the central claims of the paper better and appreciate the feedback about the initial title ("Emergent Multi-View Geometry Through Self-Distillation"). We propose instead to name the paper “Poincar3: Multi-View Self-Distillation for 3D Representation Learning”, but we are open to further changes if requested by the reviewers.

AuthorsOpenReview ↗

Second Response to Reviewer B1py

Comment [2/2]:

I'd also like to know more (intuitive or non-intuitive) explanations about why this specific task is a good measure for "emergent multi-view geometry".

We thank the reviewer for the opportunity to clarify how we measure emergent multi-view geometry. We view correspondence estimation as a key measure of 3D awareness, following Banani et al. [1] and the broader multi-view geometry literature, where point correspondences are a fundamental building block [2]. This is also consistent with recent findings that correspondence structure emerges in feed-forward reconstruction models [4, 5].

SeeSE3 [3] provides a complementary measure through motion prediction. As discussed in our response to reviewer 8xU6, we interpret this as evidence of locally consistent geometric structure rather than globally linear SE(3) representations. Together, the explicit correspondence structure in Fig. 2 and Tab. 2 and the implicit motion structure in Tab. 3 provide evidence of emergent multi-view geometry. We agree, however, that related structure is present to some extent in previous models as well, making the new title suggestion a better fit.

We refer the reviewer to the cross-scene SeeSE3 results in our response to reviewer 8xU6.

Understanding 3D and semantic performance

while empirically results, including comparison on 3D-related task and ablation study, provide evidence for a strong multi-view encoder, the analysis of why this happened is missing. For example, why is iBOT objective more preferred compared to DINOv2 in terms of 3D-related task? Why semantic understanding is much worse compared with DINO?

It has been observed in DINOv3 [6] that global and local crops can reduce patch locality, degrading dense performance. Similarly, [7] argues that aligning local and global crops can interfere with the quality of dense feature maps. In VGGT, patch-level correspondences are critical to pose estimation and geometric understanding [4]. We therefore interpret patch-level distillation as providing a better inductive bias for correspondence and geometry than objectives that place greater emphasis on global semantic invariance.

We find that this trade-off is substantially data-dependent. When trained on the same data, the DINOv2 objective achieves semantic performance comparable to our method. Since our architecture and objective do not require multi-view images, single-view data, e.g. ImageNet, could also be incorporated during training to improve semantic representations while retaining the multi-view signal. We leave this to future work.

We additionally experimented with DINOv3 initialization while keeping the rest of the training procedure unchanged. This improves semantic performance while retaining strong geometric performance. Nevertheless, we use the randomly initialized model as our main Poincar3 model to demonstrate that these geometric representations emerge from training from scratch.

Depth Anything 3 style DINO adaptation

I wonder have the authors tried Depth Anything 3 style DINO adaptation. [...] I think that would also be a reasonable baseline for "what if we directly adopt a single-view backbone into multi-view scenario".

We agree with the reviewer that this serves as a reasonable baseline and present the addtional results for Tab. 1 below:

MethodRE10K @3°RE10K @30°ScanNet++ @3°ScanNet++ @30°MegaDepth @3°MegaDepth @30°ETH3D Acc.ETH3D NCDTU Acc.DTU NC
Random init.0.017.40.011.30.035.91.240.6012.340.56
DINOv3 init.1.331.20.030.96.973.80.980.6810.870.59
DA3-style init.1.130.50.733.26.471.21.110.6312.510.59
Poincar38.368.24.567.07.474.60.810.806.730.63

As can be seen, Poincar3 also outperforms this baseline as an initialization for feed-forward reconstruction.

[1] El Banani et al., Probing the 3d awareness of visual foundation models, in CVPR 2024.

[2] Hartley and Zisserman, Multiple View Geometry in Computer Vision, Cambridge University Press, 2004.

[3] Chen et al., SeeSE3: Emergence of 3D Space in Vision Features, in NeurIPS 2026.

[4] Bratulić et al., On Geometric Understanding and Learned Priors in Feed-forward 3D. Reconstruction Models, in ECCV 2026.

[5] Starý et al., Understanding Multi-View Transformers, in ICCVW 2025.

[6] Siméoni et al., DINOv3, in TMLR 2026.

[7] Seitzer et al., Scalable Patch-Level Self-Supervised Learning, pre-print 2026.

Dear authors,

Sorry for the late reply and I appreciate your detailed response and efforts. My concerns have been addressed and I've decided to raise my score to 3.

Best, Reviewer B1py

AuthorsOpenReview ↗

Thank you Reviewer B1py

We again would like to thank Reviewer B1py for the time and effort put into the review and ultimately improving the paper through their suggestions.

Reviewer_8xU6OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Poincar3, a multi-view transformer trained from scratch by self-distillation, without 3D supervision or RGB reconstruction. An EMA teacher sees the unmasked images plus extra views of the same scene, while the student sees a masked and augmented subset, and training combines masked-patch and global distillation with anti-collapse regularization. Under fixed data, architecture, and compute, this objective outperforms RGB reconstruction for zero-shot matching on ScanNet. The scaled model improves zero-shot correspondence, frozen-backbone reconstruction transfer, and scores on a scene-specific SE(3) motion probe. The contribution is making multi-view self-distillation work from random initialization.

Critical strengths
  • A controlled objective study. With data, architecture, and compute fixed, ScanNet PCK@25 rises from 54.5 for RGB reconstruction to 66.7 with the global objective and 70.2 with extra teacher views (Table 4). This isolates what drives the gains.

  • Poincar3 learns from scratch without 3D annotations, while SelfEvo improves an existing VGGT model. Each controlled experiment uses 4×H200 for one day (§4.4), making the objective practical to study.

Critical weaknesses
  • The SE(3) interpretation is overstated. The probe predicts motion as a difference of adapter features (Eq. 5), which is exactly additive over frame triples, but the SE(3) logarithm is not additive for non-commuting motions. The probe can only fit motion locally.

  • The probe evidence is weak. The clipped mean R² of Poincar3 is only 0.098, and the reported fraction above 0.3 is 7.1% (Table 3). The error bars cover adapter seeds on 20 fixed scenes, not uncertainty over new scenes (E.3).

  • Finetuning supervision is unclear. Full finetuning is optionally regularized by distillation from a frozen VGGT-Ω teacher (E.1), but the paper does not say which rows of Table 1 use it or with what weight. Those comparisons may not be matched.

  • Layer selection is not validated. Tables 2 and 3 report the best of four encoder depths for each model (E.2, E.3), and no separate validation set is mentioned. Choosing depths on the evaluation data may inflate the reported margins.

Questions for authors
  • Which rows of Table 1 use VGGT-Ω distillation during finetuning, with what weight, and is it applied identically across all initializations?
  • Could you report unclipped per-scene R², split by stride or motion magnitude? Gains limited to small motions would narrow the SE(3) interpretation.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8xU6

We thank Reviewer 8xU6 for the careful review and valuable suggestions.

VGGT-Ω feature distillation typo

Which rows of Table 1 use VGGT-Ω distillation during finetuning, with what weight, and is it applied identically across all initializations?

As noted above in the response to Reviewer jGL1, no model uses VGGT-Ω feature distillation. There is a typo in Appendix E.1 which will be fixed in the revised manuscript.

Additivity of the SE(3) log

The SE(3) interpretation is overstated. The probe predicts motion as a difference of adapter features (Eq. 5), which is exactly additive over frame triples, but the SE(3) logarithm is not additive for non-commuting motions. The probe can only fit motion locally.

The reviewer is right that Eq. 5 is additive over triplets while the SE(3) logarithm is not and is only valid locally. We will state this limitation explicitly in the revised manuscript. This is also consistent with the motivation of the SeeSE3 protocol [1], which we follow. To quantify the regime in which the approximation becomes less accurate, we measured the non-additivity of the ground-truth translations at different temporal strides. Result is below.

Stride283260
Non-additivity in GT2.5%18%52%76%
Median pair rot.7.8°28°56°76°
As the reviewer points out, the approximation is good locally but degrades for longer trajectories.

R²

The probe evidence is weak. The clipped mean R² of Poincar3 is only 0.098, and the reported fraction above 0.3 is 7.1% (Table 3). The error bars cover adapter seeds on 20 fixed scenes, not uncertainty over new scenes (E.3).

Our original results follow the per-scene SeeSE3 [1] protocol, where one adapter is trained independently for each scene. This protocol naturally produces relatively low absolute R² values. For comparison, the original SeeSE3 paper reports that DINO obtains R² > 0 on 31% of scenes and R² > 0.3 on 3% of scenes, similar to the values we report.

To directly address the concern about generalization across scenes, we additionally evaluated the cross-scene protocol from Fig. 5 of [1]. We train a single adapter on 30 scenes and evaluate it zero-shot on 20 disjoint test scenes. The results are below.

ModelClipped R² [95% CI]Median (unclipped)Rooms R²>0Rooms R²>0.3
DINOv30.099 [0.067, 0.133]0.1074%1%
MuM0.202 [0.163, 0.240]0.2195%18%
Muskie0.224 [0.195, 0.252]0.2498%14%
Poincar30.308 [0.264, 0.350]0.3298%57%

Here the 95% CIs are percentile bootstrap intervals over the 20 held-out test rooms, addressing the concern about the original CIs, and the layer selection is chosen by 5-fold cross-validation on 30 validation scenes.

Could you report unclipped per-scene R², split by stride or motion magnitude? Gains limited to small motions would narrow the SE(3) interpretation.

In response to the reviewer’s request, we report the unclipped per-scene R² split by stride / motion magnitude below for Poincar3.

Per-scene protocol

StrideMean R²Median R²Worst roomR²>0
2+0.05+0.09−0.5680%
4+0.08+0.12−0.6380%
8+0.04+0.14−0.8274%
16−0.24+0.02−2.2653%
32−0.34−0.10−2.9031%
60−2.29−0.31−15.9525%

Cross-scene protocol

StrideMean R²Median R²Worst roomR²>0
2+0.31+0.34+0.04100%
4+0.35+0.35+0.16100%
8+0.32+0.32+0.12100%
16+0.23+0.26−0.2090%
32+0.00+0.08−0.5155%
60−0.14+0.02−1.0155%

As noted from the reviewer and expected from the local nature of the log-SE(3) approximation, the per-scene probe deteriorates at large strides.

Layer selection

Layer selection is not validated. Tables 2 and 3 report the best of four encoder depths for each model (E.2, E.3), and no separate validation set is mentioned. Choosing depths on the evaluation data may inflate the reported margins.

We agree that selecting the layer on the evaluation set can introduce optimistic bias. We originally used the best layer, following [1, 2], to avoid disadvantaged baselines whose optimal representation occurs at an intermediate depth, but we agree that held-out layer selection is a cleaner protocol. For example, using the last layer, as is often done, roughly doubles the gap between Poincar3 and the second best method in Tab. 2 (on ScanNet attention matching).

For both Tab. 2 and 3, we therefore repeated the analysis with layer selection performed on a held-out set of scenes. The resulting performance gap was essentially unchanged.

[1] Chen et al., SeeSE3: Emergence of 3D Space in Vision Features, in NeurIPS 2026.

[2] El Banani et al., Probing the 3D awareness of visual foundation models, in CVPR 2024.

AuthorsOpenReview ↗

Second Response to Reviewer 8xU6

We hope to have addressed the reviewer's concern and we are happy to further answer any remaining questions during the discussion period.

Thanks for the detailed response. The new cross-scene results, with bootstrap CIs over held-out rooms and cross-validated layer selection, make the probe evidence much more convincing. I believe this work have values for the community and tend to accept this work.

AuthorsOpenReview ↗

Thank you Reviewer 8xU6

We would again like to thank Reviewer 8xU6 for your time and effort in reviewing our work. We appreciate your positive outlook on the paper.

Reviewer_jGL1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Poincar3, a self-supervised multi-view representation learner that replaces the RGB reconstruction objective of CroCo/MuM/Muskie with DINO-style self-distillation. The backbone is a ViT-L encoder followed by a 12-layer alternating-attention multi-view decoder (~650M parameters), trained from scratch at 256² on sequences of 2–24 views. The authors identify three ingredients for stable training:

  • an image-level (CLS) distillation loss on top of iBOT-style masked-patch distillation;
  • a teacher that sees 0–12 extra views of the scene, which the student never sees;
  • full images instead of local/global crops.

The model trained on 8×H200 over 204k scenes has demonstrated strong performance against baselines, validating the core claim of the paper.

Critical strengths
  1. Controlled evidence that multi-view self-distillation trains from scratch and beats RGB reconstruction. Tab. 4 fixes the architecture, data and compute. Under these conditions, the proposed objective (V) outperforms the MuM-style RGB objective. Ablation studies demonstrate that each ingredient contributes. More importantly, the stability metrics (Fig. 16) show the teacher's features are not collapsing, which answers a key question: whether DINO-style SSL can be made multi-view without 3D labels.

  2. The proposed method achieves consistent gains across diverse probes, at lower compute than the multi-view baselines. The margins are large and appear across protocols. Fig. 9 places Poincar3 at ~1.8k A100-equivalent GPU-hours, vs. ~2.7k (Muskie) and ~4.6k (MuM). Compute therefore cannot explain the gap to the multi-view baselines, even though other confounds remain (see Weakness 1).

Critical weaknesses

W1

Major comparisons could confound the objective with data, architecture and scale; controlled evidence is narrow. (a) Table 1–3 use released checkpoints that differ in data and architecture. The Tab. 1 footnote concedes MuM/Muskie "differ substantially"; DINOv3 is single-image. (b) Tab. 4 is the only controlled study. It reports one metric (attention PCK@25), no variance, no pose or point clouds. Its largest step (VI -> VII: +13.5 / +13.0) scales compute and model size together; the ablation model's size is unspecified. (c) The data ablation compares the 3D-labeled subset (~9.8k of 204k scenes, Tab. 7) with all data, conflating label-free training with data volume.

W2

The correspondence numbers do not follow the stated best-layer protocol. Fig. 14 (feature NN, values read from the plot) disagrees with Tab. 2 for two methods:

  • DINOv3: ~40 / 56 / 77 / 57 across depths. Tab. 2 reports 56.8, which matches the last layer, not the best.
  • Poincar3: ~70 / 73 / 93 / 75. Tab. 2's 79.5 matches no depth.
Questions for authors
  1. Which depth does Tab. 2 use for DINOv3 and for Poincar3 (feature NN)?
  2. App. E.1 says finetuning is "optionally" regularized by feature distillation from a frozen VGGT-Ω. Which Tab. 1 rows use it? Distilling from a strong supervised model, if applied unequally, would confound the comparison.
  3. What are the model size for rows I–VI? Does row IV drop DINOv2's image-level loss together with local crops?
2 comments
AuthorsOpenReview ↗

Response to Reviewer jGL1

We thank Reviewer jGL1 for the thorough review and constructive feedback. We hope the clarifications and additional experiments below address the concerns raised, and we would be happy to further clarify any remaining questions during the discussion period.

Optional feature distillation typo

App. E.1 says finetuning is "optionally" regularized by feature distillation from a frozen VGGT-Ω. Which Tab. 1 rows use it?

No models use feature distillation. All comparisons use the same finetuning protocol and are supervised only with the corresponding ground-truth geometry labels. The statement in L1085 of the Appendix is an error. We experimented with VGGT-Ω distillation during development but ultimately did not use it in any reported experiment. We will remove this statement from the revised manuscript.

Correspondence estimation layer depth

Which depth does Tab. 2 use for DINOv3 and for Poincar3 (feature NN)?

The reviewer has correctly identified an inconsistency between Tab. 2 and Fig. 14 in the Appendix. The curves for DINOv3 and Poincar3 in Fig. 14 were incorrect. We will replace Fig. 14 with the correct curves in the revision. Table 2 itself does not need to change, which uses the best layers. For DINOv3 and Poincar3, this happens to be layer 18.

Tab. 4

What are the model size for rows I–VI? Does row IV drop DINOv2's image-level loss together with local crops?

The model size used for rows I–VI is a ViT-B encoder (768-dimensional embeddings) followed by a multi-view transformer of 12 layers with the same embed dimension. In total, the model has 302M parameters compared to 653M for row VII (Poincar3). As indicated by the reviewer, row IV (multi-view iBOT) has only a patch loss on global crops. We will clarify both of these points in the manuscript.

Tab. 4 is the only controlled study. It reports one metric (attention PCK@25), no variance, no pose or point clouds.

To directly address the concern that Tab. 4 only evaluates correspondence, we additionally evaluated relative pose estimation from the frozen features of all Tab. 4 models using a trained head on-top (following the protocol in Tab. 1). The results are below.

MethodRE10K @30°ScanNet++ @30°MegaDepth @30°
I: RGB reconstruction25.317.831.5
II: DINOv218.914.123.0
III: Multi-view DINOv222.615.926.7
IV: Multi-view iBOT28.121.235.4
V: +Image-level objective35.032.148.3
VI: +Teacher seeing more views41.837.254.8
VII51.948.767.5

Confounding variable in data ablation

The data ablation compares the 3D-labeled subset (~9.8k of 204k scenes, Tab. 7) with all data, conflating label-free training with data volume.

We agree that this experiment (L457) does not isolate the effect of supervision from dataset size. Its purpose is instead to quantify the practical advantage of a label-free objective. Poincar3 can exploit the full 204k-scene corpus, whereas methods requiring 3D supervision are restricted to the substantially smaller annotated subset.

In additional experiments, we found that the observed difference is largely driven by distributional coverage. The key advantage of the label-free objective is therefore its ability to leverage larger, more diverse video corpora without costly 3D annotations. We will clarify this interpretation in the revised manuscript

I appreciate the authors' effort during the rebuttal phase. My concerns have been addressed by the additional extensive experiments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Updated Manuscript

Dear reviewers and AC,

We again thank the reviewers for their thorough reviews and valuable suggestions for improvements of the paper. We have now uploaded an updated manuscript based on the revisions proposed by the reviewers. Most notably, the title is revised to "Poincar3: Multi-View Self-Distillation for 3D Representation Learning" in accordance with Reviewer B1py's points about the initial title.

To make it easy to spot the changes in the manuscript, we have marked them in red.

We are happy to engage in further discussion and are open to further revisions to the manuscript during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.