An official website of the United States government
We present an online framework for dense 3D reconstruction of indoor scenes using sequential Manhattan keyframes. We take advantage of the global geometry of indoor scenes extracted by Manhattan frames and pose optimization to enhance the accuracy and robustness of the reconstructed models. During sequential reconstruction, a Manhattan frame is extracted for each keyframe after surface normal adjustment, and used for a Manhattan keyframe-based planar alignment to initialize the surface registration while a pose graph optimization is used to refine camera poses. The final model is created by integrating the Manhattan keyframes into the unified volumetric model using refined pose estimations. Experimental results demonstrate the advantage of our geometry-based approach to reduce the cumulative registration error and overall geometric drift.
more »
« less
This content will become publicly available on June 10, 2026
GPVK-VL: Geometry-Preserving Virtual Keyframes for Visual Localization under Large Viewpoint Changes
Visual localization, the task of determining the position and orientation of a camera, typically involves three core components: offline construction of a keyframe database, efficient online keyframes retrieval, and robust local feature matching. However, significant challenges arise when there are large viewpoint disparities between the query view and the database, such as attempting localization in a corridor previously build from an opposing direction. Intuitively, this issue can be addressed by synthesizing a set of virtual keyframes that cover all viewpoints. However, existing methods for synthesizing novel views to assist localization often fail to ensure geometric accuracy under large viewpoint changes. In this paper, we introduce a confidence-aware geometric prior into 2D Gaussian splatting to ensure the geometric accuracy of the scene. Then we can render novel views through the mesh with clear structures and accurate geometry, even under significant viewpoint changes, enabling the synthesis of a comprehensive set of virtual keyframes. Incorporating this geometry-preserving virtual keyframe database into the localization pipeline significantly enhances the robustness of visual localization.
more »
« less
- Award ID(s):
- 2007613
- PAR ID:
- 10633070
- Publisher / Repository:
- IEEE
- Date Published:
- ISBN:
- 979-8-3315-4364-8
- Page Range / eLocation ID:
- 16728 to 16738
- Format(s):
- Medium: X
- Location:
- Nashville, TN, USA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Extracting good views from a large sequence of visual frames is quite difficult but a very important task across many fields. Fully automatic view selection suffers from high data redundancy and heavy computational cost, thus fails to provide a fast and intuitive visualization. In this paper we address the automatic viewpoint selection problem in the context of 3D knot deformation. After describing viewpoint selection criteria, we detail a brute-force algorithm with a minimal distance alignment method in a way to not only ensure the global best viewpoint but also present a sequence of visually continuous frames. Due to the intensive computation, we implement an efficient extraction method through parallelization. Moreover, we propose a fast and adaptive method to retrieve best viewpoints in real-time. Despite its local searching nature, it is able to generate a set of visually continuous key frames with an interactive rate. All these combine provide insights into 3D knot deformation where the critical changes of the deformation are fully represented.more » « less
-
Cross-view geo-localization aims to estimate the location of a query ground image by matching it to a reference geo-tagged aerial images database. As an extremely challenging task, its difficulties root in the drastic view changes and different capturing time between two views. Despite these difficulties, recent works achieve outstanding progress on cross-view geo-localization benchmarks. However, existing methods still suffer from poor performance on the cross-area benchmarks, in which the training and testing data are captured from two different regions. We attribute this deficiency to the lack of ability to extract the spatial configuration of visual feature layouts and models' overfitting on low-level details from the training set. In this paper, we propose GeoDTR which explicitly disentangles geometric information from raw features and learns the spatial correlations among visual features from aerial and ground pairs with a novel geometric layout extractor module. This module generates a set of geometric layout descriptors, modulating the raw features and producing high-quality latent representations. In addition, we elaborate on two categories of data augmentations, (i) Layout simulation, which varies the spatial configuration while keeping the low-level details intact. (ii) Semantic augmentation, which alters the low-level details and encourages the model to capture spatial configurations. These augmentations help to improve the performance of the cross-view geo-localization models, especially on the cross-area benchmarks. Moreover, we propose a counterfactual-based learning process to benefit the geometric layout extractor in exploring spatial information. Extensive experiments show that GeoDTR not only achieves state-of-the-art results but also significantly boosts the performance on same-area and cross-area benchmarks. Our code can be found at https://gitlab.com/vail-uvm/geodtr.more » « less
-
Umetani, N.; Wojtan, C.; Vouga, E. (Ed.)Most non-photorealistic rendering (NPR) methods for line drawing synthesis operate on a static shape. They are not tailored to process animated 3D models due to extensive per-frame parameter tuning needed to achieve the intended look and natural transition. This paper introduces a framework for interactive line drawing synthesis from animated 3D models based on a learned style space for drawing representation and interpolation. We refer to style as the relationship between stroke placement in a line drawing and its corresponding geometric properties. Starting from a given sequence of an animated 3D character, a user creates drawings for a set of keyframes. Our system embeds the raster drawings into a latent style space after they are disentangled from the underlying geometry. By traversing the latent space, our system enables a smooth transition between the input keyframes. The user may also edit, add, or remove the keyframes interactively, similar to a typical keyframe-based workflow. We implement our system with deep neural networks trained on synthetic line drawings produced by a combination of NPR methods. Our drawing-specific supervision and optimization-based embedding mechanism allow generalization from NPR line drawings to user-created drawings during run time. Experiments show that our approach generates high-quality line drawing animations while allowing interactive control of the drawing style across frames.more » « less
-
Visual scenes are often remembered as if they were observed from a different viewpoint. Some scenes are remembered as farther than they appeared, and others as closer. These memory distortions—also known as boundary extension and contraction—are strikingly consistent for a given scene, but their cause remains unknown. We tested whether these distortions can be explained by an inferential process that adjusts scene memories toward high-probability views, using viewing depth as a test case. We first carried out a large-scale analysis of depth maps of natural indoor scenes to quantify the statistical probability of views in depth. We then assessed human observers’ memory for these scenes at various depths and found that viewpoint judgments were consistently biased toward the modal depth, even when just a few seconds elapsed between viewing and reporting. Thus, scenes closer than the modal depth showed a boundary-extension bias (remembered as farther-away), and scenes farther than the modal depth showed a boundary-contraction bias (remembered as closer). By contrast, scenes at the modal depth did not elicit a consistent bias in either direction. This same pattern of results was observed in a follow-up experiment using tightly controlled stimuli from virtual environments. Together, these findings show that scene memories are biased toward statistically probable views, which may serve to increase the accuracy of noisy or incomplete scene representations.more » « less
