acceptodds
Under review as a conference paper at ICLR 2027

PreView: Active View Selection for 3D Reconstruction Without Warm-up

Abstract

Active view selection is a key component of 3D reconstruction, aiming to identify the most useful observations under a limited acquisition or reconstruction budget in practical scenarios. Effective view selection can reduce redundant data acquisition and reconstruction costs while preserving the quality of the final 3D model. However, many existing methods estimate the value of a view only after a warm-up reconstruction, where a preliminary scene representation must first be built from the initial observations. This paradigm not only introduces substantial computation before selection, but can also be unreliable when only a few initial observations are available, since the warm-up reconstruction itself may be incomplete or inaccurate. We observe that the value of a candidate view can instead be inferred from the new evidence it contributes beyond the currently observed views, without first constructing a reliable 3D scene. Based on this insight, we introduce PreView, a framework for view selection before 3D reconstruction. PreView uses feed-forward visual and geometric priors to extract features from available observations and explicitly models interactions between candidate views and the selected set to estimate each candidate's marginal contribution. This formulation supports different information regimes: when candidate images are available, PreView exploits appearance and geometric cues; otherwise, its pose-only variant performs viewpoint selection before acquisition using camera poses and the context established by the initial observations. Because candidate value is estimated directly from available observations, task preferences can also be incorporated naturally, steering the reconstruction budget between global scene quality and user-specified regions of interest. Experiments across multiple scenes show that PreView selects effective views without reconstruction warm-up, remains competitive in the strict pose-only setting, supports controllable task-aware reconstruction, and substantially reduces computation before selection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
view selection, active 3D reconstruction, 3D reconstruction, 3D Gaussian Splatting, task-aware reconstruction

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_RM9tOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PreView selects camera views for 3D reconstruction without first optimizing an intermediate NeRF or Gaussian scene. It combines feed-forward visual and geometric features with a sequential selector that prioritizes complementary views, while allowing users to favor overall reconstruction quality or specific ROIs. It supports both selection with candidate images available and selection using only known candidate poses plus the initial RGB observations.

Critical strengths

The effective view selection without reconstruction warm-up is efficient. And the paper provides controlled experiments over global vs regional reconstruction quality, and has a reasonable breadth of analysis on cross-dataset transfer, view-budget, etc.

Critical weaknesses
  1. Baselines are insufficient and the benefit of the learned proxy compared to existing proxy is not clear. a) Baselines such as farthest point or k-center sampling on camera poses and greedy coverage on the feed-forward point map are standard reconstruction-free selectors. Random selection as the only non-learned reference is insufficient, and the paper does not discuss learned-policy alternatives. b) The selector is supervised using the same weighted coverage and redundancy terms already incorporated into its inference score, without an independent reconstruction-quality target. However, Tab 4 does not compare the full method against the same sequential selection procedure with the learned correction disabled, leaving its incremental contribution unclear. This ablation is important to justify the learned proxy and is missing.
  2. The runtime comparison is unfair. Camera poses are assumed known for all methods, but baselines are charged a large sfm cost while the proposed method is charged only feature extraction (Tab.3 ~800s of colmap preparation vs ~2s for PreView). The authors should clarify which preprocessing remains necessary under the known-pose assumption and report time cost with shared preparation separated from method-specific computation, so the savings from removing reconstruction warm-up can be assessed independently.
Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Request for Feedback on Our Revised Responses

Dear Reviewer,

We sincerely apologize for disturbing you. Last week, we provided point-by-point responses to your concerns and questions, together with additional controlled experiments and methodological clarifications.In particular, we have expanded the comparisons with informative heuristic and existing learned-prior methods, and added a matched Greedy proxy-only experiment to isolate the benefits of learned score correction. We have also clarified the runtime accounting under known camera poses by separating front-end preparation, reconstruction warm-up, and view-selection costs, with additional conservative timing comparisons.

We would greatly appreciate your feedback on whether our responses have adequately addressed your concerns, or whether any questions remain. As the discussion deadline is approaching, if you have any further questions or concerns, we would be grateful if you could let us know as early as possible, allowing us sufficient time to address them before the discussion period ends.

Thank you again for your valuable and constructive feedback. Your suggestions have helped us clarify our methodological contributions and strengthen the experimental evidence supporting our method.

We sincerely appreciate your time and consideration.

Best regards,
The Authors

Thanks for the additional experiments, my major concerns are resolved.

AuthorsOpenReview ↗

Response to Reviewer RM9t

Dear Reviewer,

Thank you very much for your positive feedback and for acknowledging our additional experiments and revisions. We are pleased that our responses have addressed your major concerns. We sincerely appreciate your time, thoughtful evaluation, and constructive suggestions, which have helped us significant strengthen our work.

Best regards,
The Authors

Reviewer_Qis1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PreView, a view-selection framework that avoids the preliminary reconstruction stage required by many existing approaches. It uses pretrained DINOv2 and VGGT features to score candidate views with a preference-conditioned selector, then selects views sequentially to account for redundancy. The framework supports both RGB-aware and pose-only settings and allows users to adjust the relative importance of global reconstruction quality and region-of-interest (ROI) fidelity. Experiments examine reconstruction quality, selection overhead, and cross-dataset generalization.

Critical strengths
  1. Avoiding reconstruction warm-up is a useful objective, particularly when view selection must be repeated or performed under a limited computational budget. The reported reduction in selection overhead is an encouraging result.
  2. Supporting both RGB-aware and pose-only selection broadens its potential use. Preference conditioning also allows the selection process to prioritize an ROI rather than optimizing only scene-wide quality.
  3. The paper reports competitive reconstruction results across multiple scenes and includes ablations, runtime analysis, and cross-dataset evaluation. These provide useful evidence beyond performance on a single scene or benchmark.
Critical weaknesses
  1. Much of the evaluation focuses on methods that require preliminary reconstruction. While these comparisons demonstrate the benefit of avoiding that stage, they do not establish whether PreView is preferable to less expensive selection strategies. Beyond random sampling, systematic comparisons with pose-diversity sampling, farthest-point selection, and suitable geometry-based coverage heuristics would help clarify the value of the proposed approach.
  2. The selector is trained using handcrafted proxy utilities, but the paper does not compare it with selecting views directly using those utilities under the same sequential selection protocol. As a result, it is unclear whether learning improves reconstruction quality, mainly reduces the cost of evaluating the proxies, or provides some other benefit. This distinction is important for understanding why the learned component is needed.
  3. The evaluation assumes accurate camera poses and does not examine sensitivity to pose perturbations. In addition, selecting from an offline candidate pool is not equivalent to online acquisition, where candidate availability and pose estimates may change after each observation. The current experiments support offline selection more directly than closed-loop deployment. Either additional validation or a clearer statement of this limitation is needed.
Questions for authors
  1. How does PreView compare with simple warm-up-free strategies, such as pose-diversity or farthest-point sampling, under the same view budget and candidate pool? Please report both reconstruction quality and selection time.
  2. What happens when the handcrafted proxy utilities are used directly for sequential selection, without the learned selector? A matched comparison would clarify whether learning improves selection quality, computational efficiency, or both.
  3. How sensitive are the RGB-aware and pose-only modes to errors in camera rotation and translation? Please also clarify what information is assumed available in the pose-only experiments and which parts of the online acquisition setting remain untested.
3 comments
AuthorsOpenReview ↗

Request for Feedback on Our Revised Responses

Dear Reviewer,

We sincerely apologize for disturbing you. Last week, we provided point-by-point responses to your concerns and questions, together with additional controlled experiments and methodological clarifications.In particular, we have expanded the evaluation with stronger low-computation heuristic baselines and a matched Greedy proxy-only comparison to isolate the contribution of the learned correction. We have also conducted additional pose-perturbation experiments and clarified the input assumptions, robustness, and applicability of PreView to pre-acquisition planning and online view selection.

We would greatly appreciate your feedback on whether our responses have adequately addressed your concerns, or whether any questions remain. As the discussion deadline is approaching, if you have any further questions or concerns, we would be grateful if you could let us know as early as possible, allowing us sufficient time to address them before the discussion period ends.

Thank you again for your valuable and constructive feedback. Your suggestions have helped us clarify our methodological contributions and strengthen the experimental evidence supporting our method.

We sincerely appreciate your time and consideration.

Best regards,
The Authors

Thank you for providing the detailed rebuttal and additional experiments. The newly added baseline comparisons, matched Greedy proxy-only experiment, and pose perturbation study strengthen the empirical evaluation and provide further evidence regarding the contributions of the proposed method. Nevertheless, some limitations remain. In particular, the RGB-aware setting still incurs substantial preprocessing overhead, while the learned correction introduces additional computational costs. Moreover, although the authors have clarified that closed-loop acquisition is beyond the scope of this work, the effectiveness of PreView in practical online acquisition scenarios remains unvalidated. Overall, while the revisions address several of my initial concerns and improve the experimental evaluation, the remaining limitations constrain the practical applicability and overall contribution of the proposed approach. Therefore, I will maintain my original rating of 3 (Weak acceptance).

AuthorsOpenReview ↗

Response to Reviewer Qis1

Dear Reviewer,

Thank you very much for carefully reviewing our rebuttal and acknowledging the improvements made through our additional experiments. We sincerely appreciate your valuable feedback.

Regarding computational overhead, we acknowledge that the RGB-aware setting incurs additional front-end costs due to feature extraction from candidate images. Therefore, this setting is primarily intended for task-aware subset selection from existing image collections, aiming to reduce the subsequent reconstruction workload. Although the learned correction introduces some additional computation, our experiments demonstrate meaningful improvements in reconstruction quality, particularly in ROI regions. Meanwhile, our complete Pose-only view-selection pipeline requires only 2.171 seconds, maintaining high computational efficiency.

Regarding online closed-loop acquisition, we acknowledge that our current work has not yet been fully validated in real-world online environments. However, our newly added pose-perturbation experiments demonstrate that the view-selection process is relatively robust to camera pose errors within the evaluated range, providing preliminary evidence of its potential to tolerate pose inaccuracies in future online acquisition scenarios.

Furthermore, existing online active reconstruction methods, such as ActiveGS, typically rely on additional sensing information and continuously updated scene representations for exploration and planning. In contrast, our work focuses on efficient pre-acquisition viewpoint planning using only initial RGB observations and predefined candidate camera poses, without requiring intermediate reconstruction optimization. We believe that extending this setting to fully online closed-loop active exploration remains a valuable research direction. Accordingly, we are currently exploring how to extend our framework from offline view selection to online active exploration by incorporating real-time observation feedback and dynamic viewpoint planning. We are especially grateful for your suggestions, which have helped guide our ongoing and future research.

Thank you again for recognizing our work and for your constructive feedback. Your comments have not only helped us clarify the scope and limitations of our current approach but have also provided valuable guidance for our future research.

Best regards,
The Authors

Reviewer_h1B3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a framework for active view selection in 3D reconstruction without requiring a warm-up reconstruction. Instead of relying on an optimized scene representation to estimate view utility, the method leverages large pretrained models like VGGT and DINOv2, to extract geometric and appearance features. A preference-conditioned, set-aware transformer estimates the marginal utility of candidate views and sequentially selects informative observations while reducing redundancy. The framework supports both RGB-aware and pose-only view selection, as well as task-aware selection that balances global reconstruction quality and user-defined ROIs.

Critical strengths
  • The author proposed a solution for active view selection methods by eliminating the need for computationally expensive warm-up reconstruction.
  • The framework supports both RGB-aware and pose-only view selection, allowing informative viewpoints to be selected even when candidate images are unavailable.
Critical weaknesses
  • Limited practical justification for view selection.
    • In both settings, all candidate images are already available. It is unclear whether the additional selection overhead is worthwhile compared with directly reconstructing from more or all available images, especially given increasingly efficient 3D reconstruction methods.
  • Limited technical novelty.
    • The method mainly combines existing pretrained models with a transformer-based selector and handcrafted utility functions. The necessity of the learned selector is not clearly demonstrated, as a direct comparison with greedy selection using the handcrafted utilities is missing.
Questions for authors
  • Have the authors considered comparing the end-to-end runtime and reconstruction quality against directly using more or all available images for 3DGS reconstruction? This would help justify the practical benefits of view selection.
  • How much improvement does the learned transformer provide over the handcrafted utility function alone? Could the authors include a baseline that directly performs sequential greedy selection using the utility defined in Equation (7), without the learned correction?
3 comments
AuthorsOpenReview ↗

Request for Feedback on Our Revised Responses

Dear Reviewer,

We sincerely apologize for disturbing you. Last week, we provided point-by-point responses to your concerns and questions, together with additional controlled experiments and methodological clarifications.In particular, we have added matched-budget comparisons against larger random subsets and all-image reconstruction, including comprehensive end-to-end computation analyses. We have also introduced a matched Greedy proxy-only baseline to directly quantify the incremental contribution of the learned selector to reconstruction quality and its associated computational cost.

We would greatly appreciate your feedback on whether our responses have adequately addressed your concerns, or whether any questions remain. As the discussion deadline is approaching, if you have any further questions or concerns, we would be grateful if you could let us know as early as possible, allowing us sufficient time to address them before the discussion period ends.

Thank you again for your valuable and constructive feedback. Your suggestions have helped us clarify our methodological contributions and strengthen the experimental evidence supporting our method.

We sincerely appreciate your time and consideration.

Best regards,
The Authors

Reviewer_h1B3OpenReview ↗

Official Comment by Reviewer h1B3

Thanks for the detailed rebuttal and additional experiments. The clarification of the Pose-only and RGB-aware settings, along with the new experimental results, addresses my main concerns regarding the practical motivation and empirical validation.

I therefore update my rating to weak accept.

AuthorsOpenReview ↗

Response to Reviewer h1B3

Dear Reviewer,

Thank you very much for your positive feedback and for acknowledging our additional experiments and revisions. We are pleased that our responses have addressed your major concerns. We sincerely appreciate your time, thoughtful evaluation, and constructive suggestions, which have helped us significant strengthen our work.

Best regards,

The Authors

Reviewer_gCh5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes PreView a reconstruction warm up free active view selection method to identify useful views during a 3DGS optimization for scenarios such as constrained reconstruction budged or limited acquisition. PreView estimates each candidate's marginal contribution by leveraging pretrained VGGT and DinoV2 features. The VGGT provides geometric information where as Dino features provide region wise information for user selected areas. The mixed information regime is one of the main contributions of the paper. The paper proposes two settings for their work RGB-aware regime (candidate views are present) and pose only regime (RGB images are absent). The second contribution of the work is the computational improvement over the reconstruction based warm up based through pose only regime. The authors validate their method against random view selection followed by 3DGS optimization and reconstruction based warm up methods both with and without presence of candidate views in which their method outperforms existing baselines.

Critical strengths
  1. The problem setting is interesting and the solution seems practical. Leveraging large pretrained foundational models such as VGGT or DINO to extract geometric and appearance based features for useful view selection is intuitive.
  2. Pose-only regime provides considerable computational speed up (~2 seconds as opposed to ~486 seconds in their RGB aware regime) while the performance is comparable to the RGB aware regime.
  3. The PSNR improvement for ROI areas clearly establishes preference conditioning (Table 2).
Critical weaknesses
  1. Missing warm up reconstruction methods in related works: There are existing methods in literature include several works that do warm reconstruction free view selection for instance NeRF Director [1] or Peering into the Unknown [2]. Its important to understand how PreView is different from the prior works.

  2. RGB aware setting needs full reconstruction baseline: If the candidate images are already considered available, the argument for selection of views specially if the trade of between selection of views vs using all of them for reconstruction is not established.

  3. Comparison with Random + 3DGS: While the authors do compare against 3D GS with random view selection, relatively simple but somewhat more informed than random view selection are missing for instance farthest sampling on camera poses or omitting views with low ROI visibility.

  4. Uncertain about DINOv2's contribution in the proposed pipeline: PreView uses DINO to get appearance features for selected ROIs. However there were no ablations as to removing DINOv2 with a simple MLP or or just directly using the image features from the ROI.

  5. It is not clear what setting does Table 4 ablation use if its RGB aware or pose only. The goal of the ablation in my understanding is to isolate the two information branches contribution along side other components of their method. It only makes sense to see the ablations for pose only which provides the speed up.

[1] Xiao W, Cruz RS, Ahmedt-Aristizabal D, Salvado O, Fookes C, Lebrat L. Nerf director: Revisiting view selection in neural volume rendering. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition 2024 (pp. 20742-20751).

[2] Zhang Z, Xu F, Zhang M. Peering into the unknown: Active view selection with neural uncertainty maps for 3d reconstruction. InInternational Conference on Learning Representations 2026 Apr 20 (Vol. 2026, pp. 85781-85801).

Questions for authors

Clarifications on the following questions could be helpful:

  1. How does PreView differ from the existing baselines that avoid warm up reconstruction? Comparison with atleast one of those methods will help establish the benefits of proposed baseline.

  2. How does PreView perform against relatively easy but informed camera selection ? (as explained in weaknesses point 3)

  3. What is the trade off between selecting all views in RGB aware setting? How much quality is lost due to view selection. It would also be helpful to have some clarification on what situation warrants RGB aware scenario.

3 comments
AuthorsOpenReview ↗

Request for Feedback on Our Revised Responses

Dear Reviewer,

We sincerely apologize for disturbing you. Last week, we provided point-by-point responses to your concerns and questions, together with additional controlled experiments and methodological clarifications.In particular, we have added quantitative comparisons with NeRF Director, PUN, and stronger heuristic baselines, as well as comparisons with larger image subsets and all-image reconstruction. We have also included DINOv2 ablations and Pose-only component studies to further clarify the contributions of ROI appearance representations and individual selector components.

We would greatly appreciate your feedback on whether our responses have adequately addressed your concerns, or whether any questions remain. As the discussion deadline is approaching, if you have any further questions or concerns, we would be grateful if you could let us know as early as possible, allowing us sufficient time to address them before the discussion period ends.

Thank you again for your valuable and constructive feedback. Your suggestions have helped us clarify our methodological contributions and strengthen the experimental evidence supporting our method.

We sincerely appreciate your time and consideration.

Best regards,
The Authors

Reviewer_gCh5OpenReview ↗

Response to Author Rebuttal

I thank the authors for answering my questions in depth with explanations. Following the rebuttal my primary concerns including comparison with missing warm up reconstruction based methods and qualitative trade off of pose-only preview with full 3DGS reconstruction have been resolved. I am slightly concerned with the average ROI still being similar to 3DGS despite having ROI specific anchors but that doesn't warrant a rejection.

I update my score to weak accept.

Thanks, Reviewer gCh5

AuthorsOpenReview ↗

Response to Reviewer gCh5

Dear Reviewer,

Thank you very much for your positive feedback and for acknowledging our detailed rebuttal and additional experiments. We sincerely appreciate your recognition that our responses have addressed your major concerns.

Regarding your remaining minor concern about ROI reconstruction quality, we conducted an additional view-budget experiment to further investigate the contribution of Query-ROI Anchors. Specifically, we evaluated PreView-PoseOnly-ROI with and without Query-ROI Anchors across different view budgets, comparing both against random selection and all-image 3DGS reconstruction.

ROI reconstruction quality across different view budgets

RowViewsRandom ROI PSNRPreView-PoseOnly-ROIw/o Query-ROI Anchors
R11524.74426.81426.183
R22024.79527.63026.891
R33025.79227.95627.014
R44026.14428.00227.008
R55026.38128.01826.996
R66026.67127.97727.122
R78026.85427.46127.035
R8All27.04627.04627.046

The additional results reveal two important findings:

  • ROI-aware selection can outperform all-image reconstruction with fewer views. PreView-PoseOnly-ROI achieves its highest average ROI PSNR of 28.018 dB using 50 selected views, outperforming all-image reconstruction (27.046 dB) by 0.972 dB. (R5 Vs R8) As the selected-view budget increases further, the ROI quality gradually approaches the all-image result. (R5–R8)

  • Query-ROI Anchors contribute substantially to regional reconstruction quality. Removing these anchors consistently reduces ROI PSNR across all tested partial-view budgets. (R1–R7) At 50 views, ROI PSNR decreases from 28.018 to 26.996 dB, a reduction of 1.022 dB. (R5, Full Vs w/o Query-ROI Anchors) This further demonstrates the effectiveness of Query-ROI Anchors in guiding task-relevant view selection.

These results provide additional evidence that our ROI-specific design enables more effective allocation of limited view budgets, achieving higher regional reconstruction quality than both random selection and all-image reconstruction under certain budgets.

We sincerely appreciate your acknowledgment of our revisions and your indication that you would update your rating to Weak Accept. However, we noticed that the system currently still displays a rating of 2. If you are satisfied with our detailed point-by-point response, we would be grateful if you would consider updating your rating accordingly.

Thank you again for your valuable feedback, time, and consideration.

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.