acceptodds
Under review as a conference paper at ICLR 2027

OpenLongTail: Generative Scaling of Long-Tail Driving Data

Abstract

Scaling robust driving policies is fundamentally bottlenecked by the scarcity of edge cases in curated datasets. While the real world continuously captures these critical events, such long-tail events remain underutilized when collected from heterogeneous sources. Specifically, diverse but valuable in-the-wild long-tail videos lack the full view coverage required for training policy models, often missing multi-view poses or originating solely from monocular dash cameras. This modality gap prevents these ubiquitous observations from being converted into scalable training data for long-tail generalization. We introduce OpenLongTail, an open-source generative data engine for scaling autonomous driving policies under long-tail events. To transform heterogeneous data sources into view-aligned and temporally coherent multi-view assets that are useful for policy learning, we develop a pose-informed extrapolative view synthesis pipeline that generates the missing views. We further incorporate Plücker-ray camera conditioning and lane-weighted supervision to provide geometric guidance for missing-view synthesis while placing greater emphasis on sparse lane structures during training. By synthesizing heterogeneous long-tail data, we observe an effective improvement in closed-loop driving robustness in handling long-tail events. Using extrapolative view-synthesis and pose-recovery metrics, we validate OpenLongTail in terms of visual fidelity and ego-trajectory accuracy.

Primary area:
generative models
Keywords:
Autonomous driving, Diffusion Model, VLA

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Vkr9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OpenLongTail proposes a generative pipeline that converts monocular driving videos into synchronized multi-view training data. It combines ego-pose recovery, geometry-guided view synthesis, and joint rear-view generation with road-surface conditioning. Experiments demonstrate improved generation quality and show that the synthesized data benefits downstream closed-loop driving performance.

Critical strengths
  1. Practical significance. The paper addresses an important limitation in autonomous driving data collection. Converting abundant monocular long-tail videos into multi-view training data offers a promising way to improve data coverage.

  2. Meaningful downstream validation. The authors evaluate the generated data through closed-loop driving experiments. The results show clear improvements over nominal-data fine-tuning and approach the performance achieved with real multi-view data. This supports the practical value of the proposed pipeline.

Critical weaknesses
  1. Multi-view consistency and missing comparisons. The rear-tele view in the demos appears to exhibit substantially different motion speeds from the other rear views. This raises concerns about multi-view consistency, which has received considerable attention in recent driving generation research, including the WorldLens benchmark (Liang et al., CVPR 2026). The paper also lacks comparisons with closely related methods such as NeoVerse and Sensor2Sensor. More quantitative evaluations and explanations are needed.

  2. Unclear advantage over direct multi-view generation. Existing driving world models, such as opendwm and magicdrive-v2, can directly generate multi-view videos with explicit control over road layouts and foreground objects. What advantages does converting real front-view videos into surround-view data offer over direct multi-view synthesis for long-tail data augmentation? A comparison of downstream policy improvements would help establish the value of this approach.

Questions for authors

Fig.2 jointly generates three rear views but does not explicitly connect the front-left and rear-left views, or the front-right and rear-right views. Why are these potentially related views generated separately? Could this design limit spatial and temporal consistency across the full camera rig?

3 comments
AuthorsOpenReview ↗

Response to Vkr9 (1/2)

We thank the reviewer for the constructive feedback. Your review raises two main concerns and one question:

  • W1: Insufficient multi-view consistency evaluation and missing baseline comparisons.
  • W2: Advantages over existing direct multi-view generation methods.
  • Q1: Motivation for generating side and rear views separately.

We address these points below with additional experiments and clarifications.

W1. Multi-view consistency and missing comparisons

We clarify that the apparent motion differences across rear views are expected consequences of their distinct camera geometries. The rear-tele camera (30° FoV) looks directly backward, while the rear-left and rear-right cameras (70° FoV) capture stronger lateral parallax from roadside structures. These different camera configurations naturally produce distinct apparent-motion patterns. OpenLongTail faithfully reproduces these camera-dependent motion differences, closely matching the corresponding GT values. In contrast, GEN3C substantially overestimates rear-view motion. This result confirms that the slower rear-tele motion reflects real camera-dependent motion patterns rather than generation artifacts.

Table 1. Mean optical-flow magnitude (px/frame) across five target views on 80 held-out clips.

MethodCLCRRLRRRT
GT7.4111.676.317.565.99
OpenLongTail7.4511.076.687.316.41
GEN3C5.066.3714.7012.4512.81

Besides, we include NeoVerse [1] in an additional comparison under the original evaluation protocol (Main Table 1, Sec. 4.2), reported in Table 2. All methods are evaluated on the same 80 held-out long-tail clips, with results aggregated across five target views. Sensor2Sensor [2] is excluded because its code and checkpoints are not publicly available for comparison.

Table 2. Extrapolative view evaluation on held-out long-tail clips.

MethodPSNR ↑SSIM ↑LPIPS ↓compL1 ↓mIoU ↑Road IoU ↑Lane IoU ↑FID ↓FVD ↓
NeoVerse [1]10.680.39040.691153.760.1690.3350.02179.701147.2
Ours14.980.51980.570233.130.3400.6540.22932.92254.1

Additionally, we evaluate generation quality using seven metrics from the WorldLens [3] evaluation framework on our 80 held-out NVIDIA PAV long-tail clips. All methods are evaluated on the same clips under identical preprocessing and metric protocols, enabling a controlled comparison of object fidelity, temporal consistency, semantic stability, and cross-view coherence. Therefore, OpenLongTail still outperforms the baselines on generation quality.

Table 3. WorldLens generation evaluation on held-out long-tail clips.

MethodSubject Fid. ↑Subject Cohe. ↑Subject Cons. ↑Depth Disc. ↓Temp. Cons. ↑Sem. Cons. ↑View Cons. ↑
GT67.289.193.814.6195.594.3162.2
Ours30.386.584.215.2086.993.8131.7
GEN3C15.479.866.617.7678.787.8110.5

W2. Advantage over direct multi-view generation

Thanks. Layout-conditioned generators such as MagicDrive-V2 [4] and OpenDWM [5] synthesize all views from BEV maps, 3D boxes, and text, which monocular long-tail videos from heterogeneous sources do not provide, whereas OpenLongTail keeps the recorded front video with its real event and generates only the missing views. The front view and its front-tele crop therefore remain real observations, covering two of the four Alpamayo-R1 inputs. Under matched fine-tuning budgets, these assets improve AS from 0.717 to 0.748, approaching 0.764 with the corresponding recorded multi-view data (Main Table 2). We agree that a downstream comparison with layout-conditioned generation would be informative, and it needs BEV maps and 3D boxes for the same long-tail events, which our monocular sources lack. In summary, layout-conditioned generators create new scenes from annotations, and OpenLongTail converts recorded events into multi-camera training data.

References

[1] Yuxue Yang et al. NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos. CVPR, 2026.

[2] Jiahao Wang et al. Sensor2Sensor: Cross-Embodiment Sensor Conversion for Autonomous Driving. CVPR, 2026.

[3] Ao Liang et al. WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World. CVPR, 2026.

[4] Ruiyuan Gao et al. MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control. ICCV, 2025.

[5] OpenDWM: Open Driving World Models https://github.com/SenseTime-FVG/OpenDWM. Software repository, 2025.

AuthorsOpenReview ↗

Response to Vkr9 (2/2)

Q1. Separate generation of side and rear views

Thanks for this question. The split follows where same-time evidence exists: CL and CR overlap the front camera and are generated from its reprojection, while the rear cameras have no same-time overlap with the front camera and are generated jointly from each other and from road priors (Sec. 3.2–3.3). In the PAV rig, CL and RL (CR and RR) do not overlap at all: 0% of RL pixels fall inside the CL view and 0% of RR pixels inside the CR view, at any depth from 5 m to infinity, with a gap of about 8° between the two fields of view. A direct link would therefore attend only to unmatched tokens, and adding CL and CR to the joint rear sequence would add two video streams to a branch that already takes 486 s per clip (Appendix F.2). Both branches share the same front-view conditioning and recovered ego trajectory, which keeps the motion of the side and rear views consistent. In conclusion, the split follows the overlap structure of the rig at a practical cost.

Dear Reviewer Vkr9,

Thank you for your constructive feedback and recognition of our work's practical significance and downstream value. We truly appreciate your support.

Following your suggestions, we added a NeoVerse comparison, WorldLens consistency evaluation, and optical-flow analysis confirming that rear-view motion differences match ground truth, along with further clarifications.

As the discussion period comes to a close, we hope you may review our responses. We aim to make the paper as solid as possible and welcome any further feedback.

Thank you once again for your encouragement, time, and consideration!

Reviewer_5pXbOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes OpenLongTail, a method for generating multi-view long-tail driving data from front-view observations to improve driving policy performance in long-tail events. The work uses a geometry foundation model to extract the ego pose. The front views are generated through a combination of reprojection and diffusion to fill in missing regions, while the rear views are generated jointly using a cross-view attention model. The generated long-tail data is evaluated by using it to train a VLA-based driving policy. Experiments show that a policy trained with OpenLongTail achieves performance on par with one trained using real-world long-tail data, but did not show it goes beyond existing long-tail data's performance.

Critical strengths
  • The quality of the OpenLongtail is much better than prior works in PNSR.
  • Longtail data is one of the main challenge for self-driving and robotics, and this work propose a good way of enlarging multi-view data, the paper will be stronger if authors can showed how OpenLongtail can improve safety or coverage of longtail cases.
Critical weaknesses
  • The expeirments show that VLA trained with OpenLongTail can achieve on par perfomrance with ground truth long tail data, but didn’t show that with more longtail data, the policy can become more robust than ground truth data.
  • Suggeste to provide WaymoE2E evlauation resutls traiend with OpenLongtail, or a subsets of longtail in Nuplan showing the value of generative long-tail data beyond ground truth
  • No rear-view results for policy training to evaluate the effectiveness. Bench2Drive may provide a good evaluation suite for rear view.
Questions for authors
  • Can this work go beyond real-world caputrued long-tail events? If so, how to adapt this work for synthetic long-tail events.
  • Can OpenLongtail genreate different camera configurations for rear views? How does that impact the generation quality
4 comments
AuthorsOpenReview ↗

Response to 5pXb (1/2)

We thank the reviewer for recognizing the importance of long-tail data augmentation and the quality of our generated views. Your review raises three key weaknesses and two questions:

  • W1: Downstream value beyond ground-truth data
  • W2: Generalization to additional long-tail datasets, such as Waymo E2E.
  • W3: Downstream effectiveness of synthesized rear views.
  • Q1: Extension to synthetic long-tail events.
  • Q2: Adaptation to different rear-camera configurations.

We address these concerns below with additional experiments and clarifications.

W1. Value beyond ground truth

We thank the reviewer for raising this concern. We provide two complementary evaluations to demonstrate the downstream benefits of OpenLongTail beyond matching GT augmentation. First, OpenLongTail can outperform GT augmentation. In our six-view DiffusionDrive evaluation (Supplementary Sec. B.5), synthetic augmentation improves AS from 0.318 to 0.423, surpassing GT augmentation (0.394), while reducing collision rate from 57.2% to 41.0%. On Alpamayo-R1, OpenLongTail similarly improves AS from 0.717 to 0.748, approaching GT augmentation (0.764). Second, increasing long-tail coverage yields consistent downstream gains. Our controlled scaling study (Table 1) progressively increases synthetic data coverage from 0% to 25%, 50%, and 100% of the 1,450 NVIDIA clips. Under identical SFT budgets, AS monotonically improves from 0.717 to 0.729, 0.738, and 0.748, demonstrating that expanding long-tail coverage progressively improves policy performance without additional training budget.

Table 1. Controlled scaling of synthetic long-tail training data.

NV Syn subsetUnique long-tail clipsAS ↑
0%00.717
25%3630.729
50%7250.738
100%1,4500.748

W2. Generalization to additional long-tail datasets, such as Waymo E2E.

Thanks for your suggestion. We incorporate 2,037 unseen Waymo E2E clips converted by OpenLongTail into policy training. As shown in Main Table 2, while overall AS remains comparable (0.748 to 0.749), performance improves in challenging categories involving uncommon vehicles (0.717 to 0.765), cyclists (0.662 to 0.689), and work zones (0.935 to 0.950). We acknowledge that these improvements are category-dependent rather than uniform. As further analyzed in Supplementary Sec. C.1-C.4, the effectiveness of additional long-tail data depends on source coverage and training-data composition. These results demonstrate that OpenLongTail can incorporate additional real-world long-tail sources into multi-camera policy training.

W3. Policy results with rear views

We clarify the policy evaluation. Besides the four-camera Alpamayo-R1, we train DiffusionDrive with six cameras, including all three synthesized rear views (Appendix B.5, Table 9). With synthesized long-tail data, DiffusionDrive improves AS from 0.318 to 0.423 and reduces the collision rate from 57.2% to 41.0%, exceeding the corresponding GT recipe (0.394 AS, 44.6%). AlpaSim re-renders every camera, including the rear views, from reconstructions of real scenes, so it tests policies trained on real recordings without the CARLA rendering gap of Bench2Drive [1]. Therefore, the six-camera policy that uses our rear views improves over both nominal training and its GT recipe.

Table 2. Six-View DiffusionDrive closed-loop evaluation.

Training DataAS ↑Collision Rate ↓
Base0.23776.0%
10K Nominal0.31857.2%
10K + NV/Ext. GT0.39444.6%
10K + NV/Ext. Syn0.42341.0%

References

[1] Xiaosong Jia et al. Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving. NeurIPS Datasets and Benchmarks Track, 2024.

AuthorsOpenReview ↗

Response to 5pXb (2/2)

Q1. Synthetic long-tail events

Thanks for this question. Yes, OpenLongTail operates on monocular videos and camera geometry regardless of whether the source is real or synthetic, so front-view events from single-view driving world models can be converted into the target rig in the same way, while simulators that already render the full rig need no conversion. Converted synthetic events also inherit the physical errors of the world model that produced them, so we treat them as a complement to recorded events.

Q2. Other rear-camera configurations

Thanks for your question. Yes, OpenLongTail can adapt to different rear-camera configurations without architectural modifications. Instead of assuming a fixed camera rig, our model conditions generation on target-camera intrinsics, extrinsics, and Plücker-ray geometry. During training, these conditions are derived from the actual camera calibrations of PAV, PandaSet, and nuScenes (Supplementary Sec. E.1), exposing the model to diverse camera configurations. Our evaluation adopts a common target rig solely to enable paired comparisons with GT views and ensure downstream policy compatibility. For other rear-camera configurations, the target-camera conditions can be adjusted accordingly, with additional training data incorporated when necessary to improve coverage. While our framework supports such adaptation, quantitative generalization to previously unseen camera configurations remains to be evaluated.

Thank you for the discussion and new results, I've raised my score. Please include the above discussion/clarifications in the final paper .

Thank you for your positive feedback and for raising your score! We sincerely appreciate your constructive suggestions, which have helped us strengthen the paper. We will incorporate the additional results and clarifications into the final version.

Reviewer_X3YWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OpenLongTail, which converts heterogeneous, pose-free monocular long-tail driving videos into pose-aligned, synchronized multi-view data. The problem definition is clear and worth studying, but the experiments and methods in the paper remain insufficient.

Critical strengths
  1. To address the lack of synchronized multi-view data and poses in monocular long-tail videos, the paper builds a model to generate multi-view data; the problem is accurately identified and meaningful.
  2. The work covers metric-scale trajectory recovery, KF/RTS stabilization, front-anchored side-view extrapolation, joint rear-view generation, closed-loop policy evaluation, and open-source plans, making it fairly comprehensive.
Critical weaknesses
  1. External Waymo Syn only moves average AS from 0.748 to 0.749, while collision rate rises from 3.8% to 5.7% and complex intersections drop from 0.747 to 0.699, the “open scaling” claim is overstated.
  2. The policy uses only front + CL/CR, the rear views are not used as policy inputs, yet a substantial portion of the method is devoted to joint rear-view generation. Experiments should be added where rear views participate in policy training, or where rear-view quality is correlated with downstream gains.
  3. The front view cannot uniquely determine rear-view content, this is an ill-posed problem, and plausible generations may not reflect the actual scene. There is a lack of discussion on the risk of policy training on such generated content.
  4. The rig-specific limitation is evident: the target camera rig is fixed, and there is no evaluation across intrinsics, FOV, mounting height, or overlap topology. It remains unknown whether the method holds when the side/rear overlap relationships change, which limits “open scaling.”
  5. Section 3.2 samples each side view separately, keeps source/target as separate batch items, and restricts self-attention within each view. CL and CR can each be plausible but need not be mutually consistent.
  6. Can the authors provide a complete rig construction procedure or a unified multi-source camera alignment pipeline? This bears on how training is conducted.
  7. The paper's core claim is generative scaling, but the downstream evaluation only compares a fixed data mixture, without varying the amount of synthetic data or its proportion.
Questions for authors

Please see the weakness.

4 comments
AuthorsOpenReview ↗

Response to X3YW (1/3)

Response to Reviewer X3YW

Thank you for the constructive feedback. Your review raises seven key concerns:

  • W1: Generative scaling and mixed effects of cross-source augmentation.
  • W2: Downstream utility of synthesized rear-view observations.
  • W3: Semantic fidelity, hallucination risks, and generation reliability.
  • W4: Generalization to different camera configurations.
  • W5: Cross-view consistency of independently generated side views.
  • W6: Camera-rig construction and alignment across heterogeneous datasets.
  • W7: Controlled evidence of downstream improvements from scaling long-tail data.

We address each concern individually below, supported by additional experiments and clarifications.

W1. Generative scaling and mixed effects of cross-source augmentation

We clarify that open scaling refers to converting heterogeneous monocular sources into policy-compatible multi-view training data, instead of guaranteeing performance gains from arbitrary data mixtures. To support this claim, we have released our model checkpoints, training code, and data conversion utilities. Furthermore, our controlled scaling experiment (Table 3) demonstrates progressive AS improvements from 0.717 to 0.729, 0.738, and 0.748 as NVIDIA synthetic long-tail data increases from 0% to 25%, 50%, and 100%, under matched SFT budgets. We acknowledge that adding Waymo decreases complex-intersection AS from 0.747 to 0.699. Importantly, adding recorded Waymo data produces a comparable decline (0.743 to 0.691), indicating that this degradation is not specific to synthesized views. This is consistent with differences in long-tail scenario coverage and training-data composition: the Waymo subset contains only 52 temporary-path-reassignment examples and no human-guided-traffic-control examples, compared with 671 and 62 in NVIDIA (Appendix C.2). Meanwhile, Waymo augmentation improves performance on uncommon vehicles, cyclists, and work zones. Therefore, our open-scaling claim is supported by the released framework and controlled scaling results, while the Waymo experiment highlights the importance of relevant long-tail coverage rather than data volume alone.

W2. Downstream utility of synthesized rear-view observations

Besides the four-camera Alpamayo-R1, we train DiffusionDrive with six cameras, including all three synthesized rear views (Appendix B.5, Table 9). With synthesized long-tail data, DiffusionDrive improves AS from 0.318 to 0.423 and reduces the collision rate from 57.2% to 41.0%, exceeding the corresponding GT recipe (0.394 AS, 44.6%). In summary, the six-camera policy that uses our rear views outperforms its GT recipe.

Table 1. Six-View DiffusionDrive closed-loop evaluation.

Training DataAS ↑Collision Rate ↓
Base0.23776.0%
10K Nominal0.31857.2%
10K + NV/Ext. GT0.39444.6%
10K + NV/Ext. Syn0.42341.0%
AuthorsOpenReview ↗

Response to X3YW (2/3)

W3. Semantic fidelity, hallucination risks, and generation reliability

Thanks for pointing it out. We agree that the front view cannot determine the rear content uniquely (Sec. 5), and closed-loop training measures the resulting risk. We apply Qwen3.6-27B to filter visual artifacts and temporal/cross-view inconsistencies before policy training, achieving 96% agreement with human judgments on 100 clips (Supplementary Sec. B.1). We further evaluate semantic fidelity against GT on 80 held-out clips (Main Table 1). Our lane-weighted supervision improves rear-view Lane IoU from 0.051 to 0.229 (Main Table 4(b)). Additionally, we employ WorldLens [1] to assess object fidelity, temporal stability, semantic consistency, and cross-view coherence:

Table 2. WorldLens generation evaluation on held-out long-tail clips.

MethodSubject Fid. ↑Subject Cohe. ↑Subject Cons. ↑Depth Disc. ↓Temp. Cons. ↑Sem. Cons. ↑View Cons. ↑
GT67.289.193.814.6195.594.3162.2
OpenLongTail30.386.584.215.2086.993.8131.7
Gen3C [4]15.479.866.617.7678.787.8110.5

W4. Generalization to different camera configurations

We agree that our evaluation uses one target rig (Appendix F.3), which we fixed to match the paired ground truth and the policy input. Each target view enters through its intrinsics, extrinsics, and Plücker rays (Appendix D.1), and the generator is trained on the native cameras of PAV, PandaSet, and nuScenes (about 200K, 1.2K, and 12K clips, Table 10), whose focal lengths, mounting positions, and overlap layouts differ. On 150 held-out nuScenes val. clips with their native rig, OpenLongTail reaches PSNR 14.14 against 13.29 for Gen3C. In conclusion, the conditioning already spans varied rigs, and accuracy on unseen rigs remains a stated limitation. To further facilitate adaptation across heterogeneous data sources and camera configurations, we also open-source our model checkpoints, training code, and data conversion utilities.

W5. Cross-view consistency of independently generated side views

CL and CR share the conditions that determine their common content: they look to opposite sides of the vehicle, so an object moving between them passes through the front view, and both are conditioned on the same front-view memory, stabilized ego trajectory, and geometric reprojections (Sec. 3.2). WorldLens view consistency is 131.7 for our views and 162.2 for GT, while the best baseline is 110.5. Independent sampling therefore changes only content that CL and CR do not share.

AuthorsOpenReview ↗

Response to X3YW (3/3)

W6. Camera-rig construction and alignment across heterogeneous datasets

Thanks for your suggestion. We will add it to future versions. Our training does not require aligning different datasets to a single canonical camera rig. Instead, OpenLongTail learns camera-conditioned view generation directly from heterogeneous camera configurations across PAV, PandaSet, and nuScenes. Specifically, each training sample uses a front-view video as input, together with the target camera's intrinsics, extrinsics, and ego-motion to construct Plücker-ray conditioning (Supplementary Sec. D.1). The target camera geometry specifies the desired viewpoint, and the corresponding recorded video provides ground-truth supervision. Thus, the model learns to synthesize different target views from their respective camera configurations, without transforming all training observations into a fixed rig. At inference, we require only a monocular front-view video and the specified target-camera calibration, with ego-motion recovered from the input video. The common target rig used in our experiments is selected for GT comparison and policy compatibility.

W7. Controlled evidence of downstream improvements from scaling long-tail data

Thanks for your suggestion. We conduct an additional controlled scaling experiment using 0%, 25%, 50%, and 100% of the available 1,450 NVIDIA synthetic long-tail clips. All subsets are nested, with the same 10K nominal training pool, policy initialization, stage-specific optimization budgets, and closed-loop evaluation protocol. The result demonstrates that scaling up long-tail training data can progressively improve VLA policy performance, highlighting the value of expanding long-tail data coverage for downstream driving policies.

Table 3. Controlled scaling of synthetic long-tail training data.

NV Syn subsetUnique long-tail clipsAS ↑
0%00.717
25%3630.729
50%7250.738
100%1,4500.748

References

[1] Ao Liang et al. WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World. arXiv:2512.10958, 2025.

Dear Reviewer X3YW,

Thank you again for your time and detailed, constructive feedback. Following your suggestions, we have carefully addressed all seven concerns and conducted several additional experiments. In particular, we clarified our open-scaling claim, added a controlled data-scaling study (AS: 0.717 → 0.748), and conducted a six-camera DiffusionDrive evaluation demonstrating the downstream utility of synthesized rear views (AS: 0.318 → 0.423; collision rate: 57.2% → 41.0%). We also provided further evaluations of cross-view consistency and camera-rig generalization, along with clarifications on the Waymo augmentation results, generation reliability, and multi-source camera-rig construction.

As the discussion period is coming to a close, we sincerely hope you may have an opportunity to review our responses and additional results. We would greatly appreciate any further feedback and would be happy to address any remaining concerns.

Thank you very much for your time and consideration!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank Reviewers X3YW (R1), 5pXb (R2), and Vkr9 (R3) for their time and feedback. We appreciate the strengths you recognized:

  1. A meaningful problem: "the problem is accurately identified and meaningful" (R1).
  2. A practical way to scale multi-view data: "a good way of enlarging multi-view data" (R2).
  3. Closed-loop validation: "This supports the practical value of the proposed pipeline" (R3).

We added the following results and clarifications, detailed in the individual responses:

  • Controlled scaling study (R1, R2): AS rises from 0.717 to 0.729, 0.738, and 0.748 as NV Syn data grows from 0% to 25%, 50%, and 100% of the 1,450 longtail clips under matched SFT budgets.
  • Rear views in policy training (R1, R2): the six-camera DiffusionDrive in Appendix B.5 takes all three synthesized rear views as input; AS improves from 0.318 to 0.423 and the collision rate drops from 57.2% to 41.0%.
  • WorldLens evaluation (R1, R3): OpenLongTail leads Gen3C, the strongest baseline on visual fidelity in Main Table 1, on all seven metrics, e.g., view consistency of 131.7 vs. 110.5 (GT: 162.2).
  • Comparison with NeoVerse (R3): PSNR of 14.98 vs. 10.68 and FVD of 254.1 vs. 1147.2 on the 80 held-out clips of Main Table 1.
  • Rear-view motion analysis (R3): mean rear-tele optical flow is 6.41 px/frame for OpenLongTail vs. 5.99 for GT and 12.81 for Gen3C.
  • Evaluation on a second camera configuration (R1, R2): on the nuScenes rig, which differs from the PAV target rig, PSNR is 14.14 vs. 13.29 for Gen3C on 150 held-out val. clips.

We hope these results address the reviewers' concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.