bundle adjustment

Existing initialization-free calibration-aware bundle adjustment methods rely on pseudo-object-space errors (pOSE), which are inherently constrained by projective invariance and cannot exploit known camera intrinsics—resulting in reconstructions defined only up to a projective equivalence class, lacking metric scale and Euclidean structure. Method: We propose a novel initialization-free calibrated Structure-from-Motion (SfM) framework that incorporates known camera calibration into pairwise relative rotation estimation, then jointly optimizes rotations via rotation averaging and pOSE refinement, directly solving for a near-metric reconstruction under the similarity transformation group. Contribution/Results: Our method breaks free from projective ambiguity, explicitly preserving scale and rigid-body geometry. It converges with high probability to the global optimum even from completely random initializations. Experiments demonstrate significantly higher reconstruction accuracy and robustness compared to conventional pOSE-based approaches, establishing a new paradigm for unsupervised calibrated SfM.

Existing camera pose estimation methods suffer from low initial pose accuracy, reliance on task-specific descriptors, or the need for network fine-tuning or scene reconstruction. To address these limitations, we propose GS-SMC—the first lightweight pose optimization framework based on 3D Gaussian Splatting (3DGS). GS-SMC leverages a pre-trained 3DGS model to render multi-view images, establishes 2D–2D correspondences via feature matching, and iteratively refines poses using epipolar geometry constraints—without requiring network adaptation or scene re-reconstruction. Its modular architecture enables flexible substitution of feature extraction and matching components, eliminating dependence on specialized descriptors or networks. Evaluated on the 7-Scenes and Cambridge Landmarks benchmarks, GS-SMC reduces translational error by 53.3% and 40.7%, and rotational error by 56.9% and 53.2%, respectively—outperforming state-of-the-art approaches significantly.

This work addresses the computational inefficiency of conventional bundle adjustment in multi-camera systems, which jointly optimizes camera poses and 3D points. By leveraging the generalized camera model, the authors introduce, for the first time, a purely pose-based geometric constraint that implicitly encodes 3D points through two reference observations and their associated poses, thereby entirely eliminating 3D point parameters from the optimization. The resulting formulation optimizes only camera poses, significantly enhancing computational efficiency while maintaining or even improving pose estimation accuracy. Experimental results on both synthetic and real-world datasets demonstrate that the proposed method consistently outperforms traditional bundle adjustment in terms of speed and accuracy.

This paper addresses the relative pose estimation problem for three calibrated cameras given only four correspondences across all views. To overcome limitations of conventional methods—namely, their reliance on more correspondences or insufficient robustness—we propose a novel strategy that approximates a fifth correspondence using the centroid of the four observed points. We further introduce the first joint three-view pose estimation framework integrating a 4-point affine fundamental matrix solver, a standard 5-point relative pose solver, and a P3P solver. Geometric modeling enhances robustness against noise and outliers, while local optimization refines accuracy. Evaluated on real-world datasets, our method achieves state-of-the-art performance: the centroid-based strategy significantly outperforms pure affine approaches, striking a superior balance among accuracy, robustness, and computational efficiency, with straightforward implementation.

Traditional novel view synthesis relies on external Structure-from-Motion (SfM) tools (e.g., COLMAP), which introduce pose estimation errors and computational bottlenecks. To address this, we propose the first end-to-end framework that jointly optimizes 3D Gaussian splatting radiance fields and camera poses. Our method decouples optimization into two alternating stages—Gaussian parameter refinement and pose fine-tuning—guided by differentiable rendering and a geometric-photometric consistency constraint enforced via 3D optical flow. This design significantly suppresses projection errors, especially under large parallax and sparse-texture conditions. Evaluated on multiple standard benchmarks, our approach not only substantially outperforms existing COLMAP-free methods but also surpasses the COLMAP+3DGS baseline across both reconstruction quality and camera pose accuracy.

This work addresses the inherent scale ambiguity in feed-forward 3D reconstruction models, which hinders their ability to recover metrically accurate geometry and limits applicability in scenarios requiring metric understanding. To resolve this, the authors propose the first end-to-end trainable feed-forward framework that leverages readily available unlabeled satellite imagery as a global metric prior. By integrating satellite image retrieval, bidirectional cross-view feature interaction, and scale consistency constraints, the model recovers absolute scene scale, refines geometric structure, and estimates camera poses in a metric coordinate system—even with only coarse initial pose estimates. Evaluated on KITTI, nuScenes, and Oxford RobotCar, the method significantly improves metric depth estimation, multi-view point cloud reconstruction, and cross-view localization accuracy, while demonstrating strong generalization across datasets and geographic regions.

This study addresses the degradation in 3D Gaussian Splatting (3DGS) reconstruction quality caused by illumination variations and insufficient geometric accuracy in multi-temporal, high-resolution satellite imagery with small convergence angles. To overcome these challenges, this work introduces 2D Gaussian Splatting (2DGS) into satellite photogrammetry for the first time. The proposed method integrates an affine camera model with online camera refinement based on sparse observations, while jointly incorporating geometric shadow mapping and per-camera color correction to effectively mitigate time-varying illumination and shadow artifacts. Evaluated on the DFC2019 and IARPA2016 datasets, the approach significantly outperforms existing 3DGS-based methods, achieving an 11.93% reduction in mean absolute error and a 31% decrease in peak GPU memory consumption.

This work addresses the systematic distance distortion that arises when converting ground coordinates using planar homography with a monocular camera, where minor initial calibration errors induce distortions that grow approximately quadratically with true distance. The study derives, for the first time, an explicit analytical model linking homography perturbations to distance error, revealing their inherent quadratic relationship. Building on this insight, two lightweight correction strategies are proposed: one based on regression-fitting a quadratic error function, and another employing coordinate-wise gradient descent optimization. Evaluated on a large-scale simulation comprising over 19 million samples, the regression-based method achieves higher accuracy when the underlying model is reliable, whereas the gradient descent approach demonstrates greater robustness against poor initial calibrations.

This study addresses the challenge of unreliable manual feature matching in traditional RPC bundle adjustment for multi-temporal satellite imagery, which suffers from seasonal variations, illumination changes, and surface cover dynamics that degrade uncontrolled geometric positioning accuracy. To overcome this limitation, the authors propose an appearance-aware RPC refinement method that, for the first time, jointly leverages learned local features and global image descriptors to robustly extract season-invariant correspondences. Furthermore, the approach employs visual compatibility metrics to select optimal image pairs, thereby enhancing the match graph structure. Evaluated on a multi-season WorldView-3 dataset, the method significantly outperforms open-source baselines, achieving notably reduced geometric consistency errors and substantially improved matching efficiency across image blocks of 39–42 scenes.

This work addresses the geometric reconstruction errors introduced by existing Gaussian splatting methods that rely on perspective or affine camera approximations of Rational Polynomial Coefficient (RPC) models. We propose, for the first time, a Gaussian splatting framework natively supporting RPC cameras, directly projecting Gaussian means and covariances through the RPC model to eliminate approximation errors. Our key innovations include integrating the RPC model into the splatting pipeline, designing a Jacobian-based robust covariance projection method, and introducing ray-based depth modeling to overcome the lack of explicit depth in RPC formulations. Evaluated on the DFC2019 and IARPA2016 datasets, our approach reduces mean elevation errors by 29.6%/63.8% and 9.9%/37.9%, respectively, compared to conventional methods, significantly advancing 3D reconstruction accuracy.