Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer

• URL: http://arxiv.org/abs/2605.15828v2
• Date: Fri, 22 May 2026 08:53:58 GMT
• Title: Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer
• Abstract summary: We propose Fisher-Guided Quantization (FGQ) for feed-forward 3D reconstruction models.<n>FGQ uses the diagonal Fisher information matrix to quantify the different sensitivities across tasks, blocks, and channels.<n>Experiments across camera pose estimation, point map reconstruction, and depth estimation show that FGQ consistently outperforms state-of-the-art quantization baselines.
• Score: 26.28481512169773
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Feed-forward 3D reconstruction models, represented by Visual Geometry Grounded Transformer (VGGT), jointly predict multiple visual geometry tasks such as depth estimation, camera pose prediction, and point cloud reconstruction in a single forward pass. They have been widely adopted in 3D vision applications, but their billion-scale parameters bring substantial memory and computation overhead, posing challenges for on-device deployment. Post-Training Quantization (PTQ) is an effective technique to reduce this overhead. Existing PTQ methods for feed-forward 3D models mainly focus on handling heavy-tailed activation distributions and constructing diverse calibration datasets. However, we observe that feed-forward 3D models predict multiple geometric attributes through a shared backbone, where different transformer blocks and hidden channels contribute distinctly to each task, resulting in substantially different sensitivities to quantization errors across tasks, blocks, and channels. Consequently, treating all tasks equally over-emphasizes insensitive tasks and causes significant accuracy loss on the sensitive ones. To address this issue, we propose Fisher-Guided Quantization (FGQ) for feed-forward 3D reconstruction models. Specifically, FGQ uses the diagonal Fisher information matrix to quantify the different sensitivities across tasks, blocks, and channels, and incorporates these sensitivities into the Learnable Affine Transformation during calibration to better preserve the channels and blocks most critical to each task. Extensive experiments across camera pose estimation, point map reconstruction, and depth estimation show that FGQ consistently outperforms state-of-the-art quantization baselines on VGGT, achieving up to 39% relative improvement under the 4-bit quantization. Code is available at https://github.com/ypzhng/FGQ.

Related papers

• PointSplat: Compact Gaussian Splatting via Human-Centric Prediction [70.45838533838104]
  We propose PointSplat, a novel human-centric approach to producing 3D human representations from input views on the fly.<n>The proposed method first estimates a coarse geometric proxy and performs ray casting to prune redundant points.<n>It then employs a Point-Image Transformer to fuse appearance and geometry features, predicting Gaussian attributes in a single forward pass.
  arXiv  Detail & Related papers  (2026-06-30T17:59:05Z)
• QVGGT: Post-Training Quantized Visual Geometry Grounded Transformer [5.7977296539841845]
  We introduce QVGGT, a tailored quantization framework designed to compress Visual Geometry Grounded Transformer (VGGT)<n>QVGGT achieves near-lossless W4A16 quantization, preserving the accuracy of all 3D prediction heads while delivering 3$sim$4.9$times$ memory reduction and up to 2.8$times$ real hardware speedup over FP32.<n>Our approach makes high-fidelity 3D perception feasible on edge devices, enabling practical deployment of feed-forward 3D reconstruction models in real-world constrained environments.
  arXiv  Detail & Related papers  (2026-05-29T10:32:52Z)
• Tail-Aware Post-Training Quantization for 3D Geometry Models [58.79500829118265]
  Post-Training Quantization (PTQ) enables efficient inference without retraining.<n>PTQ fails to transfer effectively to 3D models due to intricate feature distributions and prohibitive calibration overhead.<n>We propose TAPTQ, a Tail-Aware Post-Training Quantization pipeline for 3D geometric learning.
  arXiv  Detail & Related papers  (2026-02-02T07:21:15Z)
• Visual Odometry with Transformers [68.453547770334]
  We introduce Visual odometry Transformer (VoT), which processes sequences of monocular frames by extracting features.<n>Unlike prior methods, VoT directly predicts camera motion without estimating dense geometry and relies solely on camera poses for supervision.<n>VoT scales effectively with larger datasets, benefits substantially from stronger pre-trained backbones, generalizes across diverse camera motions and calibration settings, and outperforms traditional methods while running more than 3 times faster.
  arXiv  Detail & Related papers  (2025-10-02T17:00:14Z)
• Quantized Visual Geometry Grounded Transformer [67.15451442018258]
  This paper proposes the first Quantization framework for VGGTs, namely QuantVGGT.<n>We introduce Dual-Smoothed Fine-Grained Quantization, which integrates pre-global Hadamard rotation and post-local channel smoothing.<n>We also design Noise-Filtered Diverse Sampling, which filters outliers via deep-layer statistics.
  arXiv  Detail & Related papers  (2025-09-25T15:17:11Z)
• Dens3R: A Foundation Model for 3D Geometry Prediction [44.13431776180547]
  Dens3R is a 3D foundation model designed for joint geometric dense prediction.<n>By integrating image-pair matching features with intrinsic invariance modeling, Dens3R accurately regresses multiple geometric quantities.
  arXiv  Detail & Related papers  (2025-07-22T07:22:30Z)
• VGGT: Visual Geometry Grounded Transformer [61.37669770946458]
  VGGT is a feed-forward neural network that directly infers all key 3D attributes of a scene.<n>Network achieves state-of-the-art results in multiple 3D tasks.
  arXiv  Detail & Related papers  (2025-03-14T17:59:47Z)
• Divide and Conquer: Improving Multi-Camera 3D Perception with 2D Semantic-Depth Priors and Input-Dependent Queries [30.17281824826716]
  Existing techniques often neglect the synergistic effects of semantic and depth cues, leading to classification and position estimation errors. We propose an input-aware Transformer framework that leverages Semantics and Depth as priors. Our approach involves the use of an S-D that explicitly models semantic and depth priors, thereby disentangling the learning process of object categorization and position estimation.
  arXiv  Detail & Related papers  (2024-08-13T13:51:34Z)
• IDMS: Instance Depth for Multi-scale Monocular 3D Object Detection [1.7710335706046505]
  A multi-scale perception module based on dilated convolution is designed to enhance the model's processing ability for different scale targets. By verifying the proposed algorithm on the KITTI test set and evaluation set, the experimental results show that compared with the baseline method, the proposed method improves by 5.27% in AP40 in the car category.
  arXiv  Detail & Related papers  (2022-12-03T04:02:31Z)
• Towards Model Generalization for Monocular 3D Object Detection [57.25828870799331]
  We present an effective unified camera-generalized paradigm (CGP) for Mono3D object detection. We also propose the 2D-3D geometry-consistent object scaling strategy (GCOS) to bridge the gap via an instance-level augment. Our method called DGMono3D achieves remarkable performance on all evaluated datasets and surpasses the SoTA unsupervised domain adaptation scheme.
  arXiv  Detail & Related papers  (2022-05-23T23:05:07Z)
• Geometry-aware data augmentation for monocular 3D object detection [18.67567745336633]
  This paper focuses on monocular 3D object detection, one of the essential modules in autonomous driving systems. A key challenge is that the depth recovery problem is ill-posed in monocular data. We conduct a thorough analysis to reveal how existing methods fail to robustly estimate depth when different geometry shifts occur. We convert the aforementioned manipulations into four corresponding 3D-aware data augmentation techniques.
  arXiv  Detail & Related papers  (2021-04-12T23:12:48Z)
• Reinforced Axial Refinement Network for Monocular 3D Object Detection [160.34246529816085]
  Monocular 3D object detection aims to extract the 3D position and properties of objects from a 2D input image. Conventional approaches sample 3D bounding boxes from the space and infer the relationship between the target object and each of them, however, the probability of effective samples is relatively small in the 3D space. We propose to start with an initial prediction and refine it gradually towards the ground truth, with only one 3d parameter changed in each step. This requires designing a policy which gets a reward after several steps, and thus we adopt reinforcement learning to optimize it.
  arXiv  Detail & Related papers  (2020-08-31T17:10:48Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.