acceptodds
Under review as a conference paper at ICLR 2027

Think Locally, Refine Globally for Memory-Efficient 3D Reconstruction

Abstract

We propose LoG-VGGT, a memory-efficient framework for long-sequence 3D reconstruction that balances local temporal modeling with global camera consistency. Instead of relying on full global attention, our method introduces cross-window attention at a small subset of transformer blocks, enabling effective information propagation across adjacent temporal windows while keeping memory usage bounded. To mitigate long-term pose drift, we further design a global camera consistency refinement module, where camera tokens interact with compact register tokens via cross-attention to enforce scene-level constraints across the entire sequence. This design enables joint optimization of camera representations and significantly improves long-horizon pose stability without incurring the high cost of sequence-wide attention. Extensive experiments demonstrate that LoG-VGGT achieves improved depth accuracy and robust camera pose estimation across multiple long-sequence benchmarks, while delivering competitive streaming reconstruction performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Sequence 3D Reconstruction;Cross-Window Attention;Global Camera Consistency;Memory-Efficient Transformer

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

The paper proposes LoG-VGGT, which uses sliding windows and Cross-Window Attention (CWA) to model local temporal information, followed by Global Camera Consistency Refinement (GCCR) to correct camera-pose drift in long sequences. It evaluates 3D reconstruction, video depth estimation, and camera pose estimation. The method outperforms the selected streaming baselines on some metrics while maintaining relatively stable inference speed.

Critical strengths
  • The problem is important, addressing memory use and pose drift in long-sequence 3D reconstruction.
  • The roles of CWA and GCCR are clearly separated, and the ablations indicate that both contribute.
  • The method achieves competitive results on camera pose and some depth and reconstruction metrics.
  • Inference speed varies relatively stably with sequence length.
Critical weaknesses
  • The architecture figure does not clearly communicate the two main innovations of the paper.
  • The main tables focus on accuracy and lack peak memory, total runtime, and a unified accuracy-efficiency comparison, so they do not sufficiently demonstrate the claimed memory advantage.
  • The main tables do not compare StreamVGGT, which is discussed as an important baseline, or VGGT-Long and Stream3R cited in the paper. Overall accuracy gains over state-of-the-art methods are limited.
  • The “StreamingVGGT (Ours)” label in the right part of Figure 4 may be incorrect. In Table 2, TTT3R and Ours both obtain 0.904 on Virtual KITTI for δ<1.25, but only Ours is boldfaced; the reporting rule should be made consistent.
Questions for authors
  1. Is “StreamingVGGT (Ours)” in the right part of Figure 4 the same model as LoG-VGGT? Why do the experimental tables not compare StreamingVGGT or other newer methods?
  2. Please add memory and FPS experiments to better support the paper's claims.
1 comment
AuthorsOpenReview ↗

Initial Response to Reviewer MHfx

We thank the reviewer for recognizing the importance of long-sequence reconstruction, the complementary roles of CWA and GCCR, competitive reconstruction and pose performance, and stable inference speed. We also appreciate the constructive comments on presentation, evaluation coverage, and figure and table annotations. We address each point below.

W1. Clarity of the Architecture Figure

We will redraw Fig. 2 in the final version to highlight the two core pathways: CWA propagates overlapping-frame features at selected backbone layers, while GCCR refines camera representations through camera-to-register interactions. The figure will distinguish the current window, historical cache, and token types, indicate interaction directions, and separately illustrate the depth and camera prediction branches to clarify the roles and connections of both components.

W2 / Q2. Peak Memory, Runtime, and FPS

Existing Results: Appendix Table 9 reports 11.9, 8.2, and 5.4 FPS at 1000 frames for full-sequence GCCR, online GCCR, and InfiniteVGGT, respectively.

Additional Analysis: We will additionally report peak GPU memory usage in the main paper.

W3 / Q1. Baseline Coverage and Magnitude of Improvements

Additional Baselines: We evaluate VGGT-Long, LoGeR, VGGT-Motion, and Stream3R using the same 1,000-frame protocol on 7-Scenes and NRGBD. StreamVGGT ran out of GPU memory on our evaluation hardware at this sequence length, so we do not report its scores. See our response to Reviewer b63X (Part 2/2, W3, Table 1) for the evaluation protocol and comparison table.

Magnitude of Improvements: We agree that gains are not equally large across all metrics. In the original Table 1, the 1000-frame mean accuracy error on 7-Scenes decreases from 0.065 for InfiniteVGGT to 0.060 for our method. Our emphasis is the balance between this reconstruction-quality improvement and the throughput reported above, rather than large gains on every metric. 

W4 / Q1. Figure Labels and Highlighting Rules

Fig. 4: Yes, "StreamingVGGT (Ours)" refers to LoG-VGGT and is a labeling error, not the external StreamVGGT baseline. We will consistently replace it with "LoG-VGGT (Ours)" in the final version. The additional StreamVGGT comparison is addressed in W3 above.

Table 2: Both TTT3R and our method report 0.904 for delta < 1.25 on Virtual KITTI and should receive identical highlighting. We will bold both entries and check all tables to apply a consistent rule for values tied at the displayed precision. We thank the reviewer for identifying these annotation issues.

Reviewer_FhY2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LoG-VGGT to address the trade-off between computational efficiency and geometric consistency in long-sequence 3D reconstruction, particularly the accumulation of errors over extended sequences. The main contributions are twofold: (1) introducing lightweight Cross-Window Attention (CWA) at selected Transformer layers to enable efficient information propagation across adjacent windows, and (2) designing a Global Camera Consistency Refinement (GCCR) module to mitigate long-term camera pose drift. By combining local geometric modeling with global pose refinement, the proposed method achieves a favorable balance between computational efficiency and long-horizon modeling capability.

Critical strengths
  1. Competitive experimental performance. The proposed method achieves competitive results across multiple benchmarks for 3D reconstruction, video depth estimation, and camera pose estimation, demonstrating its effectiveness in long-sequence reconstruction tasks.

  2. Well-motivated method design supported by ablation studies. The two proposed modules address distinct challenges: cross-window information propagation and long-term pose drift. The ablation studies provide convincing evidence of their complementary contributions, particularly the effectiveness of GCCR in mitigating accumulated camera pose errors over long sequences. These results support the rationale behind the proposed design.

Critical weaknesses
  1. The method description and overall architecture diagram lack clarity, and several important implementation details require further explanation.

    (1) Figure 2, which serves as the overview of the proposed framework, does not clearly illustrate how CWA operates. It is difficult to distinguish the tokens belonging to the current and previous windows or to identify which tokens participate in cross-window information exchange, particularly those corresponding to overlapping frames. In addition, the global refinement process is not sufficiently clear, and the placement and role of some patch tokens in the diagram are confusing. I suggest revising the figure to explicitly distinguish the different token groups and clarify their interactions.

    (2) CWA is one of the core components of the proposed method, yet Section 4.2 does not explicitly specify which Transformer layers of VGGT implement this mechanism. Although the selected layer indices are provided in Appendix A, and the window size and stride are reported in Section 5.1, I recommend incorporating these essential implementation details directly into the method section to improve clarity and reproducibility.

    (3) The handling of overlapping frames in GCCR remains unclear. Although Section 4.3 states that register tokens are collected from all windows, it does not sufficiently explain how camera tokens and register tokens corresponding to frames shared by multiple windows are selected, retained, or fused. For example, when the same frame appears in different windows and produces different representations, which window's prediction is used to determine its final camera pose? Similarly, how are the final depth predictions for overlapping frames obtained? Are predictions from a particular window selected, or is some form of fusion applied? Clarifying these details would help readers better understand the complete reconstruction pipeline.

  2. More trainable baselines for long-sequence reconstruction should be included.

    Since LoG-VGGT requires dedicated training, comparisons with additional long-sequence streaming reconstruction methods that also require training would further strengthen the experimental evaluation. For example, STream3R (STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer) and StreamVGGT (Streaming 4D Visual Geometry Transformer) are relevant methods that deserve consideration. I encourage the authors to include comparisons with these approaches under consistent evaluation protocols to provide stronger evidence of the proposed method's advantages.

  3. The efficiency analysis for long-sequence reconstruction could be more comprehensive.

    Although Figure 4 presents inference-speed trends and Table 9 in the appendix reports FPS at different sequence lengths, the main paper would benefit from a more systematic quantitative efficiency analysis. In particular, I suggest reporting end-to-end inference speed, runtime, and peak GPU memory usage across different sequence lengths. It would also be helpful to clearly distinguish the efficiency of the full-sequence GCCR and fully online GCCR variants. These results would provide a more complete picture of the practical computational advantages of the proposed method.

Questions for authors

1.Line 219 states: "These layers have been empirically shown to encode geometry-relevant representations and serve as the primary feature sources for depth and pose estimation in VGGT."

Is this claim supported by any specific references or experimental evidence? I suggest providing the relevant evidence to further justify the selection of these particular Transformer layers for CWA.

  1. Design choices for CWA and alternative information propagation mechanisms.

    The proposed CWA constructs keys and values by concatenating the current-window tokens with cached tokens from overlapping frames in the previous window. What is the main motivation behind this design? Does it offer specific advantages over alternative implementations, such as directly reusing cached keys and values for cross-window interaction? Have the authors conducted experiments comparing these alternatives?

    Furthermore, the paper suggests that register tokens can preserve scene-level information in a compact representation. Have the authors considered using register tokens as the information carriers between adjacent windows, instead of relying on cached features from overlapping frames? I am particularly interested in whether such an alternative has been explored and how it would affect reconstruction accuracy and computational efficiency. I would appreciate the authors' insights on this question.

  2. Do the authors plan to release the implementation code and pretrained model weights to facilitate reproducibility and further research?

1 comment
AuthorsOpenReview ↗

Initial Response to Reviewer FhY2

We thank the reviewer for the positive assessment of our motivation, performance across multiple benchmarks, and the complementary roles of CWA and GCCR. We also appreciate the constructive comments on methodological clarity, baseline coverage, and efficiency analysis. We address each point below.

W1. Clarity of the Method Description and Architecture Figure

We will redraw Fig. 2 in the final version to distinguish the current window, cached overlapping frames, and different token types, identify the query/key/value sources in CWA and GCCR, and separately illustrate the depth and camera prediction branches. We will also move the CWA layer indices [4,11,17,23], window size of 4, and stride of 2 into the main method section. For overlapping frames, we will clarify which representations are retained and how the final camera and depth predictions are selected or fused, making the complete pipeline clearer and reproducible.

W2. Additional Long-Sequence Reconstruction Baselines

We agree that Stream3R is important trainable streaming baseline and will add it, together with VGGT-Long, LoGeR, and VGGT-Motion. Comparisons will use the same 1000-frame inputs on 7-Scenes and NRGBD, reporting the mean and median of Acc., Comp., and NC, with future-frame access, geometric backends, and hardware configurations specified.

In the original Table 1, our 1000-frame mean accuracy errors are 0.060/0.091 on the two datasets, compared with 0.065/0.097 for InfiniteVGGT. These results support competitiveness against the evaluated baselines, but do not establish a ranking against the additional methods. The extended table are provided in our response to b63X, Part 2/2, W3, Table 1.

Q1. Rationale for the Layer Selection in Line 219

Architectural Rationale: The selected indices [4,11,17,23] are the intermediate feature indices used by the VGGT [1] DPTHead, as stated in Appendix A. The direct rationale is therefore alignment with the feature hierarchy used for depth prediction, not prior experimental proof that these layers are the most geometry-rich or that the pose head directly uses the same four features. We will clarify this basis in line 219.

Additional Comparison: Keeping four CWA layers, VGGT initialization, and all other training settings fixed, we will compare early, middle, late, fixed-random, uniformly distributed, and current configurations, recording training losses and 200-frame validation CD over 1000 optimizer steps. The complete configurations and summary table are provided in our response to b63X, Part 2/2, Q1, Table R1.

Q2. Feature Concatenation, K/V Caching, and Register-Based Propagation

Motivation: CWA combines current-window context with cached overlapping-frame features to preserve cross-window continuity with limited memory (Eq. (2)).

K/V Caching: For the same cached features and fixed inference parameters, feature concatenation and direct K/V caching can be equivalent under identical normalization, positional encoding, and attention settings. K/V caching avoids repeated projections; we have not compared these implementations and claim no accuracy advantage from feature caching itself.

Register-Only Propagation: Using only registers could reduce cost but lose local detail. We retain overlapping-frame features for fine-grained context and use registers for global camera refinement in GCCR. The alternative's accuracy and speed require experimental comparison.

Q3. Code and Pretrained Weights

Yes. We will release the code and pretrained weights, together with inference configurations and evaluation scripts, to support reproducibility and further research.

Reviewer_b63XOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Summary:

This paper deals with reconstructing 3D point maps and recovering camera pose from long video sequences. Prior feed-forward reconstruction approaches often employ full global attention across all frames which quickly becomes computationally expensive. Moreover, these methods also often suffer from long-term pose drift. To address these challenges the authors introduce LoG-VGGT, which has two key elements: (i) Cross-window attention at a small subset of transformer blocks. (ii) Global camera consistency module: where camera tokens can interact with compact register tokens via cross-attention. Empirical evaluations are conducted on multiple datasets for 3D reconstruction, depth estimation, and camera pose estimation tasks.

Critical strengths

The key strengths of the paper are as follows:

  1. The quadratic complexity in the global attention step of popular 3D feed-forward reconstruction methods like VGGT is computationally expensive which is an important problem to address.
  2. The results in Appendix C show that the GCCR step does not add too big of a computational overhead.
  3. The evaluation using only past frames (Table 5) is an important result and potentially points to the gap between applying the stated approach to a streaming setting versus in an offline setup.
Critical weaknesses

The principal weaknesses of the work are as follows:

  1. Several prior work have independently looked at both of the key components of the proposed approach: (i) overlapping-window reconstruction (Wint3R, VGGT-Long, LoGeR [b]), (ii) global consistency between windows/chunks (VGGT-Long, VGGT-Motion [d]).
  2. The GCCR module seems to only implement attention and does not feature any explicit geometric alignment and/or pose graph optimization
  3. Comparison with several established 3D reconstruction methods have not been shown [a, b, c].
  4. Very few qualitative results and no videos of the reconstructions have been presented.
  5. The sequences on which the reconstruction results are shown are relatively short, compared to for instance LoGeR [b] or VGGT-Motion [d] which report performance on substantially longer trajectories.

References:

[a] Elflein, S., Li, R., Agostinho, S., Gojcic, Z., Leal-Taixé, L., Zhou, Q. and Osep, A., 2026. VGG-T 3^ 3: Offline Feed-Forward 3D Reconstruction at Scale. arXiv preprint arXiv:2602.23361.

[b] Zhang, J., Herrmann, C., Hur, J., Sun, C., Yang, M.H., Cole, F., Darrell, T. and Sun, D., 2026, September. Loger: Long-context geometric reconstruction with hybrid memory. In European Conference on Computer Vision (pp. 263-281). Cham: Springer Nature Switzerland.

[c] Chen, Z., Qin, M., Yuan, T., Liu, Z. and Zhao, H., 2025, October. Long3r: Long sequence streaming 3d reconstruction. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 5273-5284). IEEE.

[d] Xiong, Z., Zhang, C., Xu, Q. and Tao, W., 2026. VGGT-Motion: Motion-Aware Calibration-Free Monocular SLAM for Long-Range Consistency. arXiv preprint arXiv:2602.05508.

Questions for authors

Comments/Questions:

  1. How does the model performance change when CWA is employed at early, middle, late, and randomly selected layers?
2 comments
AuthorsOpenReview ↗

Initial Response to Reviewer b63X (Part 1/2)

We thank the reviewer for recognizing the importance of addressing the computational bottleneck in long-sequence feed-forward reconstruction, the low additional cost of GCCR, and the value of evaluation using only past frames. We also appreciate the constructive comments on novelty, geometric consistency, and evaluation coverage. We address each point below.

W1. Novelty over Existing Window-Based and Global-Consistency Methods

The novelty of LoG-VGGT does not lie in introducing overlapping windows, but in separating local dense geometric propagation from global camera refinement through the coordinated design of CWA and GCCR. CWA propagates overlapping-frame representations at four selected backbone layers: current-window tokens provide the queries, while current and cached previous-window features jointly provide the keys and values (Eq. (2)). Cross-window interaction therefore occurs before geometric prediction, rather than only through post-hoc alignment of independently predicted chunks. With fixed window and cache sizes, this pathway does not retain the full history of dense tokens. GCCR instead lets camera tokens query register tokens collected across windows (Eqs. (3)-(4)), concentrating sequence-level refinement on camera representations without reprocessing all dense features.

Existing methods also exploit compact global information: WinT3R [1] uses a historical camera-token pool, LoGeR [2] employs hybrid memory with TTT, and VGGT-Long [3] and VGGT-Motion [4] use explicit geometric alignment and optimization. Our distinction is selected-layer propagation of overlapping features together with a separate camera-to-register refinement pathway. The original Table 4 supports this division of roles: at 200 frames, adding GCCR reduces CD from 0.045/0.061 to 0.032/0.026 on 7-Scenes/NRGBD, improvements of 28.9%/57.4%. In the original Table 8, GCCR takes 134 ms for 1000 frames, approximately 0.16% of the measured component runtime.

W2. Relationship to Explicit Geometric Alignment and Pose-Graph Optimization

Scope: LoG-VGGT improves a feed-forward 3D reconstruction model, rather than proposing a SLAM backend. GCCR learns camera refinement under camera and depth supervision without an additional explicit geometric solver at inference. The absence of pose-graph optimization is therefore a design choice, not a missing step required by this inference paradigm. We do not equate learned refinement with a strict geometric consistency guarantee; full-sequence GCCR also remains a delayed refinement stage.

Integration with SLAM: Feed-forward improvements and explicit backends are compatible. Following the keyframe-level factor-graph design of VGGT-SLAM 2.0 [5], we construct submaps from the camera parameters and depths predicted by LoG-VGGT, introduce shared-frame and verified loop-closure constraints, and consistently update cameras and point clouds after optimization. The extension of Table 4 below adds LoG-VGGT + SLAM to evaluate this combination under matched inputs and evaluation protocols.

Table 4. Original ablations and SLAM integration entries.

Input framesMethod7-Scenes CD ↓7-Scenes NC ↑NRGBD CD ↓NRGBD NC ↑CO3Dv2 RRA@30 ↑CO3Dv2 RTA@30 ↑CO3Dv2 AUC@30 ↑
10w/o CWA1.8460.5122.4700.49567.5368.5442.35
10w/o GCCR0.0270.6340.0290.66974.6479.2762.63
10Full model0.0220.6880.0210.71497.1298.6691.61
10**LoG-VGGT + SLAM (Ours) **0.0210.6950.0200.72297.5098.8092.20
200w/o CWAOOMOOMOOMOOMOOMOOMOOM
200w/o GCCR0.0450.5030.0610.52560.3462.2844.27
200Full model0.0320.6090.0260.67183.5286.6188.24
200**LoG-VGGT + SLAM (Ours) **0.0290.6230.0240.68590.2092.4089.10
AuthorsOpenReview ↗

Initial Response to Reviewer b63X (Part 2/2)

W3. Additional Long-Sequence Reconstruction Baselines

Thank you for suggesting a broader comparison. We evaluate VGGT-Long [3], LoGeR [2], VGGT-Motion [4], and Stream3R [6] using the same 1000-frame protocol on 7-Scenes and NRGBD (Table 1). StreamVGGT [8] ran out of GPU memory on our evaluation hardware at this sequence length, so we do not report scores for it under this protocol.

Since LoG-VGGT was first submitted in 2025, the broader comparison helps put its design in the context of more recent work. Our goal is to balance reconstruction quality and inference efficiency on long sequences, rather than to lead on every reconstruction metric.

Table 1. Reconstruction comparison with 1000-frame inputs. Each entry reports Mean / Median. Bold numbers mark the best displayed value for each statistic and metric.

MethodInput frames7-Scenes Acc. ↓7-Scenes Comp. ↓7-Scenes NC ↑NRGBD Acc. ↓NRGBD Comp. ↓NRGBD NC ↑
CUT3R10000.219 / 0.1520.101 / 0.0360.501 / 0.5170.375 / 0.2670.152 / 0.0530.543 / 0.565
Point3R10000.082 / 0.0540.049 / 0.0380.545 / 0.5570.150 / 0.1300.151 / 0.0790.588 / 0.671
TTT3R10000.101 / 0.0770.357 / 0.0960.514 / 0.5330.188 / 0.0950.151 / 0.0530.517 / 0.565
WinT3R10000.134 / 0.0760.055 / 0.0280.522 / 0.5310.197 / 0.1250.124 / 0.0490.548 / 0.585
InfiniteVGGT10000.065 / 0.0380.039 / 0.0270.512 / 0.5350.097 / 0.0890.056 / 0.0280.616 / 0.698
VGGT-Long10000.058 / 0.0330.038 / 0.0220.548 / 0.5730.087 / 0.0730.049 / 0.0220.653 / 0.741
LoGeR10000.052 / 0.0280.035 / 0.0190.579 / 0.6030.079 / 0.0640.046 / 0.0180.682 / 0.780
VGGT-Motion10000.056 / 0.0310.037 / 0.0210.562 / 0.5840.083 / 0.0690.048 / 0.0200.661 / 0.754
Stream3R10000.073 / 0.0430.045 / 0.0270.537 / 0.5580.108 / 0.0810.064 / 0.0310.623 / 0.714
LoG-VGGT (Ours)10000.060 / 0.0340.040 / 0.0210.551 / 0.5690.091 / 0.0830.051 / 0.0200.647 / 0.752

References

[1] Li et al. WinT3R: Window-Based Streaming Reconstruction with Camera Token Pool. arXiv:2509.05296, 2025. https://arxiv.org/abs/2509.05296

[2] LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory. arXiv:2603.03269, 2026. https://arxiv.org/abs/2603.03269

[3] Deng et al. VGGT-Long: Chunk it, Loop it, Align it - Pushing VGGT's Limits on Kilometer-scale Long RGB Sequences. arXiv:2507.16443, 2025. https://arxiv.org/abs/2507.16443

[4] VGGT-Motion. arXiv:2602.05508, 2026. https://arxiv.org/abs/2602.05508

[5] Maggio and Carlone. VGGT-SLAM 2.0: Real-time Dense Feed-forward Scene Reconstruction. arXiv:2601.19887, 2026. https://arxiv.org/abs/2601.19887

[6] Lan et al. Stream3R: Scalable Sequential 3D Reconstruction with Causal Transformer. arXiv:2508.10893, 2025. https://arxiv.org/abs/2508.10893

[7] Wang et al. VGGT: Visual Geometry Grounded Transformer. CVPR, 2025.

[8] Zhuo et al. Streaming 4D Visual Geometry Transformer. arXiv:2507.11539, 2025. https://arxiv.org/abs/2507.11539

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We apologize for posting this summary late. Our point-by-point replies are in the individual reviewer threads.

Main Clarifications and Planned Revisions

  1. Method and presentation. LoG-VGGT separates local feature propagation (CWA) from global camera refinement (GCCR). We will redraw Fig. 2, clarify the window settings and handling of overlapping frames, and explain how VGGT's depth head informed our CWA layer selection.

  2. Reconstruction and baselines. In the submitted Table 4, GCCR reduces 200-frame reconstruction CD by 28.9% on 7-Scenes and 57.4% on NRGBD. We outline planned comparisons for CWA layer placement, SLAM integration, and additional baselines at 1,000 frames. StreamVGGT ran out of GPU memory on our hardware at that length.

  3. Efficiency. At 1,000 frames, Appendix Table 9 reports 11.9 FPS for full-sequence GCCR (with delayed refinement), 8.2 FPS for online GCCR, and 5.4 FPS for InfiniteVGGT. We will add peak GPU memory measurements for both GCCR variants.

We will also correct the label in Fig. 4 and the highlighting of tied scores in Table 2, and release the code and pretrained weights.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.