- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a training-free acceleration of feed-forward 3D reconstruction models (VGGT family) by giving them fewer input views. For each candidate image, it computes its low-resolution DINO features then uses farthest-point sampling (FPS) in this feature space to pick a fixed fraction of views (15%), starting from the first frame. The frozen VGGT model runs only on the selected views. On 7 Scenes, NRGBD, ScanNet-50 and Tanks and Temples, using 15% of the views keeps reconstruction and pose quality close to the full input. On ScanNet-50 it gives 13–24× speedups and 63–74% lower peak memory, including the selection cost.
Critical strengths
- Simple, training-free and plug-and-play. The selector works without any training, and reduces runtime by 13–24× and peak memory by 63–74% on ScanNet. The reported times include the selection cost, which makes the efficiency numbers honest. This matters because reconstruction from long image sets is a real bottleneck for these models.
Critical weaknesses
-
Uniform sampling gives almost the same results as the proposed FPS on DINO features, which weakens the main contribution. In the paper's own ablation (Tab. 5, 6), Uniform is roughly the same as the proposed sampling strategy, while also being faster. The main tables (Tab. 1–4) only compare Full vs. + Ours and do not include Uniform at all. The speed and memory gains come from using fewer views, which any subsampling gives. So the paper does not show that its selection rule, the main contribution, really matters.
-
All test sets are densely captured, mostly indoor videos, where uniform sampling already works well. Except for Tanks and Temples, all datasets (7 Scenes, NRGBD, ScanNet-50) are dense indoor videos with very redundant frames. For a small room like those in NRGBD or ScanNet-50, there is no need to reconstruct the geometry from 500 or 1000 views. In my experience, the more overlapping views are fed to VGGT, the larger the chance that the model outputs layered artifacts in the geometry, which makes the metrics worse. This explains why selecting far fewer views gives better quality here, both in numbers and visually. So the lesson here is that these datasets just need far fewer views. It does not explain why FPS on DINO features is needed. A proper demonstration would run the method on challenging, very long, large-scale inputs (e.g., the sequences used in VGGT-Long or TTT3R) and show that the proposed selection beats both the baselines and simple uniform sampling.
-
The method solves a simpler task, so the speedup comparison is not fair. The method only predicts depth and poses for the selected views, and pose is evaluated only on these views. As noted above, for small rooms, feeding fewer views even makes it easier for the model to get good results. In contrast, Full inference and acceleration methods like FastVGGT and SparseVGGT give outputs for all N frames. Many downstream tasks need every frame results. So the speedup compares two different outputs, and I do not think this comparison is fair.
Questions for authors
-
Could author add Uniform, Random and fixed-stride rows to Tables 1–4?
-
Could the method work well on very long, large-scale inputs (e.g., the sequences used in VGGT-Long or TTT3R) and other settings where uniform sampling should fail (non-uniform camera speed, unordered image collections). Does FPS beat Uniform there?
-
All three backbones are VGGT-based: original VGGT, FastVGGT and SparseVGGT. So the claim that view selection applies "across backbones" (abstract) is tested on VGGT model family only. Why not other feed-forward models with global cross-view attention, such as Depth Anything 3 or Pi3?