acceptodds
Withdrawn from ICLR 2027

PruneBA: Efficient Visual Geometry Initialization with Chunked Progressive Bundle Adjustment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward visual geometry foundation models provide camera hypotheses for image collections that challenge feature-based Structure-from-Motion (SfM). Turning these hypotheses into supported sparse geometry requires efficient neural inference and effective geometric refinement. We introduce PruneBA, a training-free framework combining shared random patch-token pruning in the Visual Geometry Grounded Transformer (VGGT) with chunked progressive bundle adjustment (BA). Independently inferred overlapping windows undergo local triangulation and BA; a deterministic centrality criterion selects camera estimates for global re-triangulation and BA. Across seven ground–UAV, wide-baseline, and indoor evaluation cases, unpruned progressive BA retains all 157 case-level RGB cameras, compared with 58 for classical incremental SfM, and lowers internal mean reprojection error relative to direct global BA in every case. On Bus Ground–UAV 30, progressive refinement increases point yield from 96 to 298; adding 75% pruning yields 364 points at 1.364 pixels while retaining all 30 cameras. Synthetic ten-frame inference achieves a 2.63× speedup with random weights and 2.20× with pretrained weights, with lower GPU allocation in both studies. Exact attention-pair counts and finite-population sampling identities characterize the computational and spatial effects of pruning, while a centrality surrogate formalizes camera selection. These results establish sparse neural aggregation and staged geometric refinement as a practical combination for visual geometry initialization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, structure from motion, bundle adjustment, visual geometry transformers, token pruning, efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ti9QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents PruneBA, a training-free framework that aims to make VGGT-based 3D reconstruction more efficient. The main idea is to combine random visual token pruning with a progressive bundle adjustment pipeline. Specifically, the method keeps the same subset of patch tokens across images within each window to reduce computation, and then processes overlapping image windows independently using local triangulation and bundle adjustment. The locally refined camera estimates are selected using a simple centrality rule and passed to a final global refinement stage. The authors evaluate the method on seven reconstruction cases and report improved camera coverage over traditional SfM and lower reprojection errors compared to direct global BA. They also show that pruning 75% of visual tokens can speed up pretrained VGGT inference by around 2.2×. Overall, the paper explores a practical way to combine lightweight neural inference with classical geometric optimization.

Critical strengths

I think the paper tackles a practical problem, as VGGT can be quite expensive when processing multiple images, and improving its efficiency without additional training is definitely useful. The overall framework is fairly simple and easy to follow, and combining token pruning with progressive bundle adjustment seems like a reasonable design. I also like that the authors test the method on different types of scenes, including wide-baseline and indoor settings, rather than focusing on just one scenario. The reported 2.2× inference speedup with pretrained weights is encouraging, and the paper provides fairly detailed implementation and timing results. Overall, the approach is straightforward and has some practical value.

Critical weaknesses

My main concern is whether the reported improvements really translate into better 3D reconstruction. The paper mainly relies on internal reprojection error, but this error is computed on different sets of reconstructed points for each method. For example, on Office14, the number of points drops from 1,186 to just 164, so the lower error does not necessarily mean the reconstruction is more accurate. I also find the handling of independent windows a bit unclear. Since each window performs its own inference and BA, the resulting cameras may have different coordinate systems or scales, but the method directly combines selected camera estimates without explicit alignment. It is not clear why the centrality-based selection would solve this issue. Another concern is the limited novelty of the method, since random token pruning and local-to-global BA are both fairly standard ideas, and the paper does not provide enough ablations to show why these particular choices work better than simpler alternatives. The pruning quality results are also only reported on one real scene, without random-seed analysis. Finally, while the neural inference speedup looks promising, the paper does not really demonstrate an end-to-end efficiency improvement under comparable conditions. I think stronger geometric accuracy evaluations and more convincing comparisons with existing efficient VGGT methods are needed to support the main claims.

Questions for authors

I have a few questions about the method and evaluation. First, since each window is processed independently, how does the method handle possible differences in coordinate systems and scales when combining the selected camera poses? I'm not fully convinced that the centrality rule alone is enough here. Second, the paper mainly reports internal reprojection error, which is measured on different reconstructed points. Could the authors provide camera pose errors against ground truth to better demonstrate the actual reconstruction quality? Third, why is shared random token pruning preferred over other simple sampling strategies, and how sensitive are the results to different random seeds? Finally, could the authors provide a more direct comparison of end-to-end runtime and reconstruction accuracy with existing efficient VGGT methods? I think these clarifications would help better understand where the improvements really come from.

Reviewer_4vijOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The author proposed PruneBA which is a modification to the existing foundation 3D model pipeline that takes in all images in one short. The author proposed to first prune some tokens in a relatively trivial manner, and then perform local BA in local window of sequences of images first before doing it globally for all images. Empirically, their method leads to faster speed in predicting the camera locations.

Critical strengths

The paper is relatively easy to understand and the proposed method in itself makes sense. Some empirical performance gain is demonstrated.

Evaluation was performed on relatively comprehensive set of datasets. However, the evaluation baseline comparisons may not be the most suitable, see weakness.

Critical weaknesses

The author’s specific focus on BA is really a subset of the 3D foundation model prediction problem a lot of other works have already considered. And there is no lack of work for the long sequence setting the authors are trying to deal with. Some of them include VGGT-Long, VGGT-SLAM, GLUEMAP among many others.

[1] VGGT-Long: Chunk it, Loop it, Align it -- Pushing VGGT's Limits on Kilometer-scale Long RGB Sequences [2] VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold [3] Global Structure-from-Motion Meets Feedforward Reconstruction

The work, as presented, does not have specific novel contributions compared to the existing works in this field. Without the empirical evaluation against them, it is hard to tell if there are improvement conceptually and/or empirically compared to existing works.

Questions for authors

How does the proposed method compared with other variants of VGGT that do the same task?

Are there anything in particular make this works setting different from those other variants that lead to the authors omitting such comparisons?

Reviewer_Dvb4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

PruneBA is a training-free pipeline for initializing camera poses and sparse geometry. It randomly prunes VGGT patch tokens, runs VGGT on overlapping windows with local triangulation and BA, picks one camera per image by a centrality rule, and then runs global BA. The method is evaluated on seven small cases with 9–63 images each.

Critical strengths
  • The target setting, ground/UAV and wide-baseline captures where incremental SfM fails, is practically relevant.
  • The method is simple and training-free.
Critical weaknesses
  1. Insufficient experiments.
    • There is no external accuracy metric. GT poses for the 7-Scenes cases are available but not used (App. B.1).
    • The only metric is internal reprojection error, which is not comparable across models that keep different point sets. On Office14 (Table 2), progressive BA keeps 164 points against 1,186 for direct global BA.
    • Most models have a mean track length of about 2.0, so almost every point is observed in only two views.
    • Only seven small cases are evaluated. Pruning is tested on real images in a single case with a single run.
    • In the indoor cases, depth PNGs were fed into VGGT (Sec. 4.1).
    • Key baselines are missing: VGGT+BA, VGGSfM, MASt3R-SfM, and GLOMAP.
  2. Poor writing. The paper reads like an experiment log, with recurring phrases such as "saved-model" and "historical depth inputs". The elementary propositions in the appendix are overstated as contributions in the abstract.
Questions for authors
  1. What is the pose accuracy against the 7-Scenes GT?
  2. What are the indoor results without the depth PNG inputs?

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

Related papers

Discussion (0)

Sign in to comment.