- 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.