The study investigates which layers of feed‑forward geometry models—specifically VGGT, π³, and VGGT‑Ω—are essential for preserving camera poses and dense 3D structure. By pruning 3,018 configurations and evaluating seven metrics across indoor and outdoor datasets, the authors identify two redundancy regions (early and late) and show that combined deletions degrade performance additively, enabling more efficient pruning. They also demonstrate that CKA can serve as a cheaper proxy for interval degradation, and that closed‑form linear calibration can recover accuracy without retraining, reducing aggregator parameters by up to 44% while maintaining comparable performance.
By Fengyi Zhang, Holger Caesar, Xiangyu Sun, Zheng Zhang, Zi Huang, Yadan Luo
arXiv:2610.07883v1 Announce Type: new
Abstract: A correctly reconstructed distant point appears uncertain even when a frozen 3D model processes equivalent inputs because its output frame rotates frac...
By Sammam Mahdi, Fariha Binta Salim, Rakin Bin Rabbani, Aniqua Nusrat Zereen
The paper argues that internal self-consistency checks cannot guarantee the accuracy of photogrammetric reconstructions, a limitation that is structural rather than a tuning issue. It introduces a track‑leakage‑free hold‑out protocol that withholds a deterministic subset of images and tests each against only 3D points supported by at least two retained images, ensuring no view is evaluated against the structure it helped create. Experiments on diverse datasets show that while the protocol is well‑posed, it saturates at a confidence score of 1.00 and fails to detect coherent distortion, missing large errors that can reach over 100 m.
whyItMatters":"The study highlights that hold‑out self‑validation scores, increasingly used as quality evidence for metric deliverables, may be misleading and cannot replace external survey validation."
By Behnam Asadi
arXiv:2609.09507v1 Announce Type: new
Abstract: Learned image matching has experienced significant progress in recent years, culminating in robust and accurate matchers such as RoMa, whose robustness...
By David Nordstr\"om, Xinyue Zhang, Thibaut Loiseau, Vincent Lepetit, Fredrik Kahl
GeoCond is a lightweight reliability adapter that enhances frozen feed‑forward 3D reconstruction backbones by reading their predicted geometry to produce pose‑level uncertainty and a refinement gate. It can be trained using permutation‑orbit variance, ground‑truth pose error, or cycle residuals from unlabeled pose graphs, and at inference requires only a single backbone pass plus a small MLP. On the VGGT backbone, GeoCond reduces out‑of‑distribution AUSE from 0.32 to 0.20, transfers zero‑shot to outdoor extreme‑view scenes, and prevents collapse from uniform bundle adjustment, while also enabling gated refinement, pose‑graph weighting, calibration, curation, and capture decisions.
By David Ahmedt-Aristizabal, Mohammad Ali Armin, Russell Tsuchida, Lars Petersson
Feed-forward 3D foundation models such as VGGT predict cameras, depth, and point maps in a single pass, but can fail silently under low overlap, low parallax, and extreme relative rotation. Stratified...
The paper demonstrates that a targeted adversarial perturbation can reduce a vision‑language model’s training loss to near zero for a fixed target caption, yet the same model, when generating freely, still produces the correct description. This phenomenon, termed the train/inference gap, is traced to a single autoregressive step where the target token’s rank is fixed across all images, and further analysis shows that the language decoder, rather than the visual encoder, determines whether the corrupted signal is amplified or suppressed. The study uses a controlled two‑stage PGD attack on Qwen2.5‑VL‑7B‑Instruct and evaluates the effect on 200 held‑out COCO images, revealing that adversarial robustness in autoregressive VLMs largely depends on the language decoder’s prior.
whyItMatters":"The findings suggest that defenses and faithfulness evaluations for deployed vision‑language models should focus on the language decoder rather than the visual encoder, as the former is the key determinant of robustness to adversarial perturbations."
By Arun Josephraj Arokiaraj, Zekun Wu, Adriano Koshiyama
arXiv:2609.13308v1 Announce Type: cross
Abstract: A companion evaluation found that naming the target part in a manipulation prompt increased action accuracy by 0.32-0.63 across eight vision-language...
By Sarthak Sattigeri
The paper audits the confidence outputs of seven feed‑forward 3D reconstruction backbones across 13 datasets, evaluating four properties: error ranking, average error‑to‑uncertainty ratio, slope of this ratio, and coverage of the implied error distribution. While confidence ranks errors well, the decoded uncertainty is consistently too small—off by at least 2.4× on median cases—and worsens with higher confidence. A post‑hoc power‑law fit per backbone‑dataset pair improves all four metrics at the dataset level, reducing the median error by 1.35×, but fails to correct coverage for many held‑out scenes, indicating the models lack the correct error scale and distribution shape.
By Nanxing Nick Deng, Qing Cheng, Niclas Zeller, Daniel Cremers
arXiv:2610.23039v1 Announce Type: new
Abstract: Feed-forward Gaussian models have made novel view synthesis almost instantaneous. Feed-forward methods predict Gaussians directly, either in two stages...
By Viruthshaan Vinasirajan, Ranga Rodrigo
OccludeBench is a manually curated benchmark that tests vision‑language models (VLMs) on their ability to recover fully occluded objects from remaining visual evidence. The benchmark distinguishes direct cases—requiring recovery from a specific physical cue such as a shadow or reflection—from inferential cases, which rely on broader scene context. Across five VLMs, accuracy averages 42.7 % on direct cases versus 62.7 % on inferential cases, a 20‑point gap, while humans achieve 63.4 % and 68.5 % respectively with no significant gap. The study introduces OccludeTrace, a four‑stage diagnostic that isolates the failure point, revealing that models struggle most when recovering the relevant object from the full image projection. Interventions such as inpaint‑then‑confront reduce the gap to a statistically non‑significant difference for models with image‑generation capabilities.
By Jasin Cekinmez, Addison J. Wu, Ryo Mitsuhashi, Linrong Cai, Xingyu Fu, Zhuang Liu
arXiv:2606. 13156v2 Announce Type: replace-cross Abstract: Letting a vision-language model (VLM) think longer at test time has driven much recent progress.
By Animesh Tripathy, Aswanth Krishnan