Loading...
Organizing Committee
Politecnico di Milano
University of Missouri
University of Texas at Austin
Brown University
University of Minnesota
Abstract
3D reconstruction of scenes from image data is a classic problem in computer vision. Interest in this problem is at an all-time high, thanks to a wealth of applications which impact society in powerful ways---augmented reality, autonomous driving, cultural heritage, localization, mapping, and space exploration, to name a few. Some of these applications demand reliability of 3D reconstructions, e.g., for human safety, which translates to difficult and unsolved problems in mathematics. Other applications require efficiency at an ever-increasing scale, which calls for novelty in computational methods. This workshop will bring together experts working at an exciting crossroads where engineering, geometry, and optimization impact 3D reconstruction.
Loading...
Loading participant list in background...
Workshop Schedule
Monday, July 13, 2026
8:30 - 8:50 AM EDT
Check In
11th Floor Collaborative Space
8:50 - 9:00 AM EDT
Welcome
11th Floor Lecture Hall
Brendan Hassett, ICERM/Brown University
9:00 - 9:45 AM EDT
When are two pinhole camera images the same?
11th Floor Lecture Hall
Speaker
Rekha Thomas, University of Washington
Session Chair
Timothy Duff, University of Missouri
Abstract
In this talk, I'll address the problem of when two images of two sets of world points in projective cameras can be the same up to a projective transformation. I will also describe the locus of camera centers that allow this coincidence. The results rely on classical invariant theory and the algebraic geometry of moduli spaces of ordered points in the projective plane.
10:00 - 10:30 AM EDT
Coffee Break
11th Floor Collaborative Space
10:30 - 11:15 AM EDT
Single-View Rolling-Shutter SfM
11th Floor Lecture Hall
Speaker
Petr Hrubý, KTH
Session Chair
Timothy Duff, University of Missouri
Abstract
Rolling-shutter (RS) cameras are ubiquitous, but RS SfM (structure-from-motion) has not been fully solved yet. This work suggests an approach to remedy this: We characterize RS single-view geometry of observed world points or lines. Exploiting this geometry, we describe which motion and scene parameters can be recovered from a single RS image and systematically derive minimal reconstruction problems. We evaluate several representative cases with proof-of-concept solvers, highlighting both feasibility and practical limitations.
11:30 AM - 12:15 PM EDT
Critical Two-View Configurations via Homaloidal Parametrizations
11th Floor Lecture Hall
Speaker
Luca Magri, Politecnico di Milano
Session Chair
Timothy Duff, University of Missouri
Abstract
Fundamental matrix estimation becomes ill-posed in degenerate two-view configurations. While practical tests exist for planar scenes and pure rotation, detecting more general critical surfaces—such as hyperboloids, cones, or cylinders containing the 3D points and camera centers—remains challenging. In this talk, I will present a new degeneracy test based on the geometry of homaloidal nets of conics. This viewpoint allows us to model the image-to-image relation induced by a critical surface as a quadratic transformation, which can be estimated linearly from point correspondences. Unlike previous methods, the test does not require a pre-computed fundamental matrix, making it more stable in precisely the cases where epipolar estimation is ill-posed.
12:30 - 2:30 PM EDT
Lunch/Free Time
2:30 - 3:15 PM EDT
Balancing Novel View Synthesis and Geometric Accuracy in Gaussian and non-Gaussian Splatting
11th Floor Lecture Hall
Speaker
Philippos Mordohai, Stevens Institute of Technology
Session Chair
Timothy Duff, University of Missouri
Abstract
Traditionally, 3D reconstruction has been addressed by Multi-View Stereo (MVS), which has achieved great geometric accuracy, but has treated "texture-mapping" the inferred surfaces as an afterthought. By focusing solely on geometric criteria, 3D representations extracted using MVS methods have yielded unsatisfactory Novel View Synthesis (NVS) results. On the other hand, recent methods based on radiance fields have generated compelling synthesized views, but their representations is of lower geometric quality because they rely primarily, or even exclusively, on photometric losses. In this talk, I will present ongoing research on 3D representation and reconstruction mechanisms that aim to infer 3D models that are both geometrically accurate and able to synthesize novel views well. I will present two approaches based on explicit radiance field representations and inference-time optimization that introduce ways for the primitives to interact in 3D, complementing the current viewpoint-based losses. We named the first approach Radiant Triangle Soup (RTS) due to the representation which is in the form of translucent triangle primitives. The key novelty of RTS is the introduction of soft connectivity forces between triangles during optimization, encouraging explicit, but soft, surface continuity in 3D. The second approach is named TVGS for Tensor-Voting based Gaussian Splatting. TVGS also enables direct communication among primitives to enhance the geometric structures they form in 3D. This is accomplished by Tensor Voting, which was originally designed to infer structures from noisy inputs and has been adapted here to provide supervision during inference-time optimization, leading to more accurate scene geometry. The 3D losses introduced in both approaches do not require rendering and can be combined with essentially all losses previously reported in the literature.
3:30 - 4:00 PM EDT
Coffee Break
11th Floor Collaborative Space
4:00 - 4:45 PM EDT
Bundle-theoretic methods in geometric reconstruction
11th Floor Lecture Hall
Speaker
Jose Perea, Northeastern University
Session Chair
Timothy Duff, University of Missouri
Abstract
Many tasks in data science, including geometric reconstruction, distributed dimensionality reduction, data alignment and coordinatization, can be phrased as local-to-global inference problems with well-defined topological obstructions. The goal of this talk is to describe some of these constructions, and the emergent theoretical/algorithmic challenges in geometric and topological data analysis.
5:00 - 6:30 PM EDT
Reception
11th Floor Collaborative Space
Tuesday, July 14, 2026
9:00 - 9:45 AM EDT
Rolling Shutter Poly Nets Decompositions
11th Floor Lecture Hall
Speaker
Tomas Pajdla, Czech Technical University in Prague
Session Chair
Federica Arrigoni, Politecnico di Milano
10:00 - 10:30 AM EDT
Coffee Break
11th Floor Collaborative Space
10:30 - 11:15 AM EDT
Avoiding Collapse in Translation Averaging: A Triangle Puzzle Approach
11th Floor Lecture Hall
Speaker
Yunpeng Shi, Princeton University
Session Chair
Federica Arrigoni, Politecnico di Milano
Abstract
Translation averaging seeks to recover global camera locations from noisy pairwise translation directions and is a central component of Structure-from-Motion. The problem is challenging because direction measurements contain no distance information. As a result, standard formulations can be ill-conditioned, collapse many cameras to a single point, or be misled by structured outliers that support a coherent but incorrect reconstruction.
I will present TriP, a local-to-global framework based on a simple geometric observation: the directions along a nondegenerate triangle determine the relative lengths of its three edges. TriP treats each triangle as a local reconstruction piece, known only up to an unknown scale, and stitches these pieces together by requiring overlapping triangles to agree on their shared edges. In the logarithmic domain, this becomes a robust one-dimensional synchronization problem on a graph whose nodes represent triangles. The synchronized edge-length estimates are then used to recover the global camera locations.
This intermediate metric layer has two important benefits. It rules out zero-scale collapse by construction and provides higher-order redundancy against adversarial, cycle-consistent, and other structured corruptions. I will discuss deterministic exact-recovery guarantees under bounded adversarial corruption, including a nonvanishing corruption tolerance on complete observation graphs. I will also present experiments on synthetic and real Structure-from-Motion datasets, where TriP improves robustness and accuracy over existing translation averaging methods while scaling to graphs with more than one million cameras.
Joint work with Zhekai Fan, Wanze Li, and Jinxin Wang.
11:30 AM - 12:15 PM EDT
Geometric Algorithms for Robust and Scalable 3D Registration
11th Floor Lecture Hall
Speaker
Rene Vidal, University of Pennsylvania
Session Chair
Federica Arrigoni, Politecnico di Milano
Abstract
Rigid registration has been a fundamental problem in 3D reconstruction for decades. Yet existing methods often rely on known point correspondences or fail to scale to millions of points and extreme outlier rates. In this talk, I will present two recent advances that address these challenges. First, I will present ARCS, an algorithm for jointly estimating rotations and unknown correspondences between partially overlapping point sets, which combines efficient geometric search with robust consensus maximization and Riemannian optimization. Second, I will present TEAR, a robust and scalable registration algorithm based on a truncated entry-wise absolute residual objective that decomposes the original six-dimensional optimization into lower-dimensional subproblems, which can be solved globally by efficient branch-and-bound. Together, these methods illustrate how geometric reasoning leads to mathematically principled registration algorithms that can handle million-scale point sets with severe corruption.
12:30 - 2:30 PM EDT
Lunch/Free Time
2:30 - 3:15 PM EDT
Optimization in the Era of Feed-Forward Structure-from-Motion
11th Floor Lecture Hall
Speaker
Linfei Pan, ETH Zurich
Session Chair
Federica Arrigoni, Politecnico di Milano
Abstract
Optimization-based structure-from-motion (SfM) has been developed intensively for decades and has produced solid, well-understood results. Recently, end-to-end, feed-forward reconstruction pipelines have attracted intense interest, showing promising results on inputs where classical SfM tends to fail. This surge of interest makes it timely to revisit classical optimization-based methods and reconsider their role in the new era. In this talk, I will first revisit a global SfM method, GLOMAP, and then discuss how optimization and feed-forward models can be combined into a modern SfM pipeline, GLUEMAP, that draws on the strengths of both.
3:30 - 4:00 PM EDT
Coffee Break
11th Floor Collaborative Space
4:00 - 4:45 PM EDT
Fast and Robust 4D Reconstruction via Point Track Processing and Attention Chains
11th Floor Lecture Hall
Speaker
Yoni Kasten, Technion
Session Chair
Federica Arrigoni, Politecnico di Milano
Abstract
Recovering dynamic point clouds and 4D meshes from casual videos is a central goal in computer vision, but existing optimization-based methods remain slow and computationally expensive. This talk presents two recent approaches that accelerate 4D reconstruction by leveraging neural correspondences and the underlying mathematical structures of real motion.
The first approach, TracksTo4D, is a learning-based method that infers dynamic point clouds and cameras via a single feed-forward pass. Trained without 3D supervision by minimizing 2D reprojection errors, the network operates directly over sparse point tracks. By modeling movement with a low-rank approximation that explicitly decomposes scenes into static and dynamic components, our method achieves robust temporal point cloud and camera reconstruction at a fraction of the computational cost of traditional optimization techniques.
The second approach introduces a training-free 4D mesh generation framework driven by Spatio-Temporal Attention Chains. By treating attention matrices within a generative backbone as soft transition maps, our method establishes dense correspondences across space and time without explicit feature matching. To animate the mesh, sparse landmarks are tracked through these chains and their trajectories are robustly smoothed, with the resulting motion propagated via geodesic rigid skinning. Solving a weighted Procrustes alignment to compute local rigid transformations ensures volume preservation and generates topologically consistent 4D meshes in seconds. Ultimately, our approach scales reliably to longer sequences and enables camera estimation alongside 4D tracking.
Wednesday, July 15, 2026
9:00 - 9:45 AM EDT
Large Scale Geo-Imagery Data Collection with Applications to 3D Reconstruction and Beyond
11th Floor Lecture Hall
Speaker
Luc Vincent, Nexar
Session Chair
Joe Kileel, University of Texas at Austin
10:00 - 10:30 AM EDT
Coffee Break
11th Floor Collaborative Space
10:30 - 11:15 AM EDT
Orthogonal matrix retrieval with spatial consensus for 3D unknown-view tomography
11th Floor Lecture Hall
Speaker
Jane Zhao, University of Illinois at Urbana-Champaign
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
Unknown-view tomography (UVT) reconstructs a 3D density map from its 2D projections at unknown random orientations. A line of work starting with Kam (1980) employs the method of moments (MoM) to solve UVT in the frequency domain, assuming that the orientations are uniformly distributed. Such approaches bypass the estimation of individual viewing orientations that traditional MAP estimation requires. Kam's autocorrelation method leads to an underdetermined system, with missing orthogonal matrices. In this talk, we extend the previous orthogonal matrix retrieval (OMR) methods to jointly recover the density map and the orthogonal matrices such that spatial constraints, such as nonnegativity, can be easily incorporated. This is enabled by the closed-form expressions for spatial autocorrelation features. Numerical results show that the new approach is more robust and performs significantly better than the previous state-of-the-art OMR approach in the typical low-SNR scenario of 3D UVT.
11:30 AM - 12:15 PM EDT
Motion Averaging in 3D Computer Vision
11th Floor Lecture Hall
Speaker
Venu Madhav Govindu, Indian Institute of Science
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
Introduced around two decades ago, motion averaging provides a conceptually novel pathway to the classic problem of 3D reconstruction. It has since matured into a principled framework for camera motion estimation that hits the sweet spot of theoretical richness and empirical utility. In this talk I will delineate the motivation and development of motion averaging and its role in global SfM. I will illustrate the arguments by examining the paradigmatic method of rotation averaging that affords an interesting interplay of mathematical ideas (geometry, Lie groups, graph theory) and related considerations that arise in practice (optimization, robust statistics). I will also briefly describe the allied problem of translation averaging and outline some of its salient attributes and challenges.
12:25 - 12:30 PM EDT
Group Photo (Immediately After Talk)
11th Floor Lecture Hall
12:30 - 2:00 PM EDT
Lunch/Free Time
2:00 - 2:45 PM EDT
Rate-distortion for compressive signal parameterizations, and a finite-sample lens on when, why, and how diffusion posterior samplers fail
11th Floor Lecture Hall
Speaker
Sara Fridovich-Keil, Georgia Institute of Technology
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
This is a two-part talk. In the first part of the talk, I'll introduce a simple but effective framework for bounding the compression error of representing a signal with a compressive parameterization, such as an implicit neural representation. These bounds do not require access to the ground truth signal, are proven to hold (in error norm) for several representative signal parameterization models, are efficient to compute, and are empirically informative about local (pixel-level) compression error.
In the second part of the talk, I'll turn to posterior sampling using diffusion models, which are a powerful way to use data-driven priors in inverse problems. While posterior sampling approaches such as DPS often work well, they can fail unexpectedly and catastrophically. We introduce a finite-sample lens on posterior sampling that allows us to probe when, why, and how popular approaches such as DPS fail, including allowing us to observe how hallucinations arise during posterior sampling.
3:00 - 3:45 PM EDT
Map Synchronization in the Era of Generative Models
11th Floor Lecture Hall
Speaker
Qixing Huang, The University of Texas at Austin
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
Computing consistent maps (i.e., dense correspondences or rigid transformations) among a collection of objects is a fundamental problem across many applications areas including 3D reconstruction from images and 3D scans. This talk will begin with surveying classical optimization-based techniques. We will then make a connection between map synchronization and generative models. The talk will focus on a series of work on map synchronization by learning shape generative models from both deformable shapes and man-made shapes. We will conclude the talk by discussing future directions.
3:45 - 4:00 PM EDT
Coffee Break
11th Floor Collaborative Space
4:00 - 4:10 PM EDT
Subspace method of moments for ab initio reconstruction of cryo-EM
Lightning Talks - 11th Floor Lecture Hall
Speaker
Yuguan Wang, Princeton University
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
Cryo-electron microscopy (cryo-EM) is a widely used technique for recovering the 3-D structure of biological molecules from a large number of experimentally generated noisy 2-D tomographic projection images of the 3-D structure, taken from unknown viewing angles. Through computationally intensive algorithms, these observed images are processed to reconstruct the 3-D structures. Many popular computational methods rely on estimating the unknown angles as part of the reconstruction process, which becomes particularly challenging at low signal-to-noise ratios. The method of moments (MoM) offers an alternative approach that circumvents the estimation of viewing orientations of individual projection images by instead estimating the underlying distribution of the viewing angles, and is robust to noise given sufficiently many images. However, the method of moments typically entails computing higher-order moments of the projection images, incurring significant computational and memory costs. To mitigate this, we propose a new approach called the subspace method of moments (SubspaceMoM), which compresses the first three moments using data-driven low-rank tensor techniques as well as expansion into a suitable function basis. The compressed moments can be efficiently computed from the set of projection images using numerical quadrature and can be employed to jointly reconstruct the 3-D structure and the distribution of viewing orientations. We illustrate the practical applicability of SubspaceMoM through numerical experiments using up to the third-order moment on synthetic datasets with a simplified cryo-EM image formation model, which significantly improves the reconstruction resolution compared to previous MoM approaches.
4:10 - 4:20 PM EDT
Computing smooth surfaces from projections of geodesics
Lightning Talks - 11th Floor Lecture Hall
Speaker
Tom Gilat, Technion
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
I will explain the challenge of reconstructing a 3d network of geodesics on an underlying unknown "flattest" possible surface, when given images consisting of projections of these geodesics. This part will not be technical but I will emphasize the motivation and significance of this task in sketch-to-3d applications and CAD interfaces. I will very briefly describe how technically one can approach the following task of computing the smooth surface itself from the network of geodesics in 3d, i.e. filling the holes in the network of curves by "flattest" patches with the geodesic constraint satisfied.
4:20 - 4:30 PM EDT
Fundamental Constraints on Camera Centers Arising from 3D-2D Matches of Points and Lines
Lightning Talks - 11th Floor Lecture Hall
Speaker
Taci Kucukpinar, University of Missouri
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
We investigate combinations of 3D-2D matches of points and lines for a single calibrated camera that imply a nontrivial constraint on the camera's position. Specifically, we address the minimal cases for which n point matches and m line matches constrain the camera center to lie on an algebraic surface in R³. For two points (m,n) = (0,2), the constraint is well-known to be a self-intersecting torus. We complete the classification problem by deriving explicit, general equations for the cases (m,n) ∈ {(2,0), (1,1)} for two lines and one point-one line, respectively. We also report preliminary experiments investigating the suitability of these constraints for pose estimation tasks.
4:30 - 4:40 PM EDT
Linear Fundamental Matrix Estimation from 7 or 5 Points
Lightning Talks - 11th Floor Lecture Hall
Speaker
Juan Mogollon, University of Missouri
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
We revisit the problem of estimating the fundamental matrix of a pair of perspective cameras, a cornerstone of geometric computer vision. As is well-known, linear solvers require at least 8 point correspondences, whereas nonlinear minimal solvers require just 7 in the uncalibrated case or 5 in the calibrated case. In this paper, we consider a special case of the 7-point problem where 5 of the points are configured to lie on two lines, which has previously been shown to have a unique solution. As a theoretical contribution, we offer an analysis of how this uniqueness manifests in the standard 7-point algorithm. On a practical level, we provide the first practical linear solver for the minimal problem associated to this special configuration. Additionally, we evaluate a heuristic 5-point fundamental matrix solver based on the construction of virtual midpoints. When combined with early non-minimal fitting, the runtime and accuracy of our solver is competitive with the state-of-the-art (SoTA) on multiple benchmarks.
4:40 - 4:50 PM EDT
Convex relaxation approaches for high-dimensional optimal transport
Lightning Talks - 11th Floor Lecture Hall
Speaker
Tianyun Tang, The University of Chicago
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
Optimal transport (OT) is a powerful tool in mathematics and data science but faces severe computational and statistical challenges in high dimensions. We propose convex relaxation approaches based on marginal and cluster moment relaxations that exploit locality in the distributions. These methods approximate high-dimensional couplings using low-order marginals and sparse moment statistics, yielding semidefinite programs that provide lower bounds on the OT cost with greatly reduced complexity. For Gaus- sian measures with sparse correlations, we prove an exponential convergence rate for the cluster moment relaxation and an improved statistical error bound. We also establish approximation error bounds for the marginal relaxation when the reference measures are local perturbations of mean-field measures. In addition, we demonstrate how to extract transport maps from our relaxations, offering a simpler and interpretable al- ternative to neural networks in generative modeling. Extensive numerical experiments demonstrate strong empirical performance across a range of distributions. Our results suggest that convex relaxations can provide a promising path for dimension reduction in high-dimensional OT.
4:50 - 5:00 PM EDT
Convex Relaxations for the Optimization of Markov Processes
Lightning Talks - 11th Floor Lecture Hall
Speaker
HONGYI ZHANG, The University of Chicago
Session Chair
Joe Kileel, University of Texas at Austin
Abstract
We study the problem of optimizing Markov processes that interpolate between two prescribed probability distributions while minimizing a given cost. The main computational challenge is the curse of dimensionality: in high-dimensional state spaces, representing the full distribution is intractable. To address this, we reformulate the problem in terms of sequential couplings and develop convex relaxations based on local marginals and cluster moments. These relaxations exploit locality and sparse interaction structure, provide computable lower bounds, and recover low-order statistics of the intermediate laws. We identify dynamic optimal transport as a special case of our Markov process optimization problem and develop a procedure for recovering the underlying Benamou–Brenier dynamics from the relaxed solution. We also show that the procedure extends to more general Markov processes and illustrate it with a constrained process between Ising models.
Thursday, July 16, 2026
9:00 - 9:45 AM EDT
Robust Rotation Averaging in Global Structure-from-Motion
11th Floor Lecture Hall
Speaker
Dániel Baráth, ETH Zürich
Session Chair
Gilad Lerman, University of Minnesota
Abstract
Global SfM reconstructs an entire scene in a single shot, and its accuracy hinges on a deceptively simple step: recovering each camera's absolute orientation from a graph of noisy pairwise measurements. I will start from first principles - the geometry of SO(3), how relative rotations arise from feature matches and essential matrices, and why averaging rotations is fundamentally different from averaging numbers. The second half of the talk centres on three of our recent contributions. First, a certifiably optimal solver for 3D rotation averaging that recasts the problem as discrete inference on a voxelized SO(3) grid. Second, an analogous treatment of 1D rotation averaging, where the lower-dimensional structure admits sharper guarantees and faster solvers. Third, pose graph filtering: identifying and removing corrupted edges before averaging, so that downstream solvers operate on a clean subgraph.
10:00 - 10:30 AM EDT
Coffee Break
11th Floor Collaborative Space
10:30 - 11:15 AM EDT
Certifiably Correct Machine Perception
11th Floor Lecture Hall
Speaker
David Rosen, Northeastern University
Session Chair
Gilad Lerman, University of Minnesota
Abstract
Many perception and state estimation tasks in robotics and computer vision are naturally formalized as high-dimensional optimization problems that are known to be computationally hard (NP-hard) to solve in general; this class includes (for example) the fundamental problems of rotation averaging (in computer vision) and simultaneous localization and mapping (in robotics), among many others. Nevertheless, in this talk we present a class of certifiably correct estimation algorithms that are provably capable of efficiently recovering verifiably globally optimal solutions in many practical settings. Our approach is based upon convex relaxation: first, we develop convex (semidefinite) relaxations whose minimizers we prove provide exact, globally optimal solutions to the original estimation task for sufficiently small measurement noise; next, we describe specialized, structure-exploiting semidefinite optimization algorithms that enable even large-scale instances of these relaxations to be solved efficiently in practice. We illustrate the design of this class of methods on several fundamental 3D spatial estimation tasks (including rotation averaging, pose-graph optimization, and extrinsic camera calibration), demonstrating that our approach enables the recovery of globally optimal solutions to large-scale reconstruction problems involving tens to hundreds of thousands of variables in a matter of seconds.
11:30 AM - 12:15 PM EDT
Averaging Algorithms for Essential and Fundamental Matrices and Robust Projective Factorization
11th Floor Lecture Hall
Speaker
Ronen Basri, Weizmann Institute of Science
Session Chair
Gilad Lerman, University of Minnesota
Abstract
Global approaches to Structure from Motion (SfM) seek to recover camera poses and 3D structure in multiview settings by simultaneously handling information from all input images. In this talk, I will present two approaches to global SfM. The first approach relies on a complete algebraic characterization of essential/fundamental matrices in multiview settings. We use this characterization to introduce novel manifold averaging algorithms that jointly solve for camera positions, orientations, and calibration parameters. The second approach uses projective factorization, which we implement using a deep network with a suitable equivariant architecture.
12:30 - 2:30 PM EDT
Lunch/Free Time
2:30 - 3:15 PM EDT
Shedding Light on 3D Cameras
11th Floor Lecture Hall
Speaker
Mohit Gupta, University of Wisconsin-Madison
Session Chair
Gilad Lerman, University of Minnesota
Abstract
The advent (and commoditization) of low-cost 3D cameras is revolutionizing many application domains, including robotics, autonomous navigation, human computer interfaces, and recently even consumer devices such as cell-phones. Most modern 3D cameras (e.g., LiDAR) are active; they consist of a light source that emits coded light into the scene, i.e., its intensity is modulated over space, and/or time. The performance of these cameras is determined by their illumination coding functions.
I will talk about our work on developing a coding theory of active 3D cameras. This theory, for the first time, abstracts several seemingly different 3D camera designs into a common, geometrically intuitive space. Based on this theory, we design novel 3D cameras that achieve up to an order of magnitude higher performance as compared to the current state-of-the art. I will also briefly talk about our work toward developing `All-Weather’ 3D cameras that can operate in extreme real-world conditions, including outdoors (e.g., a robot navigating outdoors in bright sunlight and poor weather), under multi-camera interference (e.g., multiple robots navigating in a shared space such as a warehouse), and handle optically challenging objects such as shiny metal (e.g., for an industrial robot sorting machine parts).
3:30 - 4:00 PM EDT
Coffee Break
11th Floor Collaborative Space
4:00 - 4:45 PM EDT
Monocular Depth Priors in Minimal Camera Pose Estimation
11th Floor Lecture Hall
Speaker
Zuzana Kúkelová, Czech Technical University
Session Chair
Gilad Lerman, University of Minnesota
Abstract
Minimal solvers are a fundamental building block of modern geometric vision pipelines, enabling efficient and robust estimation of camera motion and calibration parameters within RANSAC-based frameworks. Recent advances in monocular depth estimation have opened new opportunities for incorporating depth priors into classical geometric estimation problems. The central question is whether such depth information can be leveraged to improve camera pose estimation beyond what is achievable using image correspondences alone.
In this talk, I will present our recent work on depth-aware minimal solvers for camera pose estimation. The main focus will be on RePoseD, a framework for relative pose estimation from image correspondences augmented with monocular depth predictions. Since modern depth estimators typically produce depths only up to unknown scale or affine transformations, we develop minimal solvers that jointly estimate camera motion together with the unknown depth parameters. The resulting methods apply to calibrated and uncalibrated camera settings and provide improvements in efficiency and robustness compared to existing depth-aware approaches. Extensive experiments demonstrate when and how monocular depth estimates can improve relative pose estimation in practice.
Beyond relative pose estimation, I will briefly discuss other examples where approximate depth information simplifies geometric estimation, including camera pose estimation from object bounding boxes. I will conclude by arguing that geometry is not only a consumer of learned depth priors but can also serve as a tool for evaluating them. When depth predictions are used in downstream tasks such as pose estimation and structure-from-motion, the performance of the resulting geometric pipelines offers a natural benchmark that complements traditional pixel-wise depth evaluation and may better reflect practical utility.
The talk highlights how imperfect but informative depth estimates can be integrated into minimal solvers, leading to faster, more robust, and more practical geometric estimation methods while also providing new perspectives on the evaluation of learned depth representations.
Friday, July 17, 2026
9:00 - 9:45 AM EDT
A differential-geometric approach to fast homotopy continuation in 3D computer vision
11th Floor Lecture Hall
Speaker
Ricardo Fabbri, Rio de Janeiro State University at New Fribourg
Session Chair
Benjamin Kimia, Brown University
Abstract
Recovering the 3D geometry of points, curves, surfaces from uncalibrated perspective views is a core problem in computer vision, lying at the intersection of projective differential geometry and algebraic geometry. Leveraging the highly optimized framework MINUS initiated at ICERM since 2019, modern methods from numerical algebraic geometry (NAG) comprising dynamical systems on complex varieties — namely fast homotopy continuation (HC) and numerical monodromy — have recently made globally convergent multivariate polynomial solving feasible for real-time applications. Because symbolic-oriented solvers have been largely shown to break down under the weight of nonlinear complexity, fast HC has become the mandatory engine for safety-critical vision tasks like autonomous driving, with many problems currently on the verge of real-time tractability. We argue that further unlocking extreme sequential computational speeds is possible, but requires a radical departure from generic local path-tracking heuristics. This talk describes a mathematical foundation suitable for fast homotopy continuation and 3D vision. We start with the modeling of continuous structure from motion involving jets of families of curves observed from video using tools from fiber bundles, connections, gauge theory, Riemannian metrics, and singularity-aware variants. We then build these differential-geometric tools up to the tracking of complex solutions to families of polynomials, using Hermitian bundles and metrics as guides to fast algorithms. Numerical path tracking in HC is framed as generalized ODE flows of sections defined by various connections in a bundle. While certain connections naturally recover the usual Davidenko flow, others minimize an action and define geodesics or general minimal paths. Optimizing the connection itself to minimize an action elevates optimal homotopy generation to a PDE problem. We show how the approach integrates the classical condition length of Shub into a geometrically coherent framework, and how Hermitian, Kähler and Hamiltonian structure relate to reducing sequential complexity bounds.
10:00 - 10:45 AM EDT
Going Beyond Pairs: Higher-Order Synchronization via Trifocal and Quadrifocal Constraints
11th Floor Lecture Hall
Speaker
Daniel Miao, University of Minnesota Twin Cities
Session Chair
Benjamin Kimia, Brown University
Abstract
In global Structure-from-Motion(SfM) methods, many works have explored synchronization based on pairwise measurements. However, few have investigated the possibility of incorporating higher order measurements into the SfM pipeline, even though higher order measurements contain more complex geometric information. In this talk, we introduce a higher-order synchronization framework that leverages triplet and quadruplet view geometries. We establish new theoretical constraints on a set of trifocal and quadrifocal tensors, expressed as low multilinear rank tensors. We then present two practical frameworks—TrifocalSync and QuadSync for trifocal and quadrifocal tensor synchronization. Our methods show the potential of higher-order synchronization by improving location estimation quality on benchmark SfM datasets.
11:00 - 11:15 AM EDT
Coffee Break
11th Floor Collaborative Space
11:15 AM - 12:00 PM EDT
Combining Geometry and Learning in Vision
11th Floor Lecture Hall
Speaker
Viktor Larsson, Lund University
Session Chair
Benjamin Kimia, Brown University
Abstract
Geometric estimation has long relied on principled models and optimization, but many real-world settings challenge explicit modeling assumptions. In this talk, I will present recent work from my group and collaborators on integrating machine learning into geometric estimation pipelines. The core theme is that learning can complement, rather than replace, classical geometry: for example, it can model noise and uncertainty that are difficult to parameterize analytically, and it can strengthen estimation in ambiguous or underconstrained cases by introducing learned priors about scene structure. I will discuss examples spanning room layout estimation, visual localization, camera pose estimation, structure-from-motion, and image matching, and I will close with a preview of ongoing directions.
Request Reimbursement
This section is for general purposes only and does not indicate that all attendees receive funding. Please refer to your personalized invitation to review your offer.
- ORCID iD
- As this program is funded by the National Science Foundation (NSF), ICERM is required to collect your ORCID iD if you are receiving funding to attend this program. Be sure to add your ORCID iD to your Cube profile as soon as possible to avoid delaying your reimbursement.
- Acceptable Costs
- 1 roundtrip between your home institute and ICERM
- Flights on U.S. or E.U. airlines – economy class to either Providence airport (PVD) or Boston airport (BOS)
- Ground Transportation to and from airports and ICERM.
- Unacceptable Costs
- Flights on non-U.S. or non-E.U. airlines
- Flights on U.K. airlines
- Seats in economy plus, business class, or first class
- Change ticket fees of any kind
- Multi-use bus passes
- Meals or incidentals
- Advance Approval Required
- Personal car travel to ICERM from outside New England
- Multiple-destination plane ticket; does not include layovers to reach ICERM
- Arriving or departing from ICERM more than a day before or day after the program
- Multiple trips to ICERM
- Rental car to/from ICERM
- Flights on a Swiss, Japanese, or Australian airline
- Arriving or departing from airport other than PVD/BOS or home institution’s local airport
- 2 one-way plane tickets to create a roundtrip (often purchased from Expedia, Orbitz, etc.)
- Travel Maximum Contributions
- New England: $350
- Other contiguous US: $850
- Asia & Oceania: $2,000
- All other locations: $1,500
- Note these rates were updated in Spring 2023 and supersede any prior invitation rates.
- Reimbursement Requests
Request Reimbursement with Cube
Refer to the back of your ID badge for more information. Checklists are available at the front desk and in the Reimbursement section of Cube.
- Reimbursement Tips
- Scanned original receipts are required for all expenses
- Airfare receipt must show full itinerary and payment
- ICERM does not offer per diem or meal reimbursement
- Allowable mileage is reimbursed at prevailing IRS Business Rate and trip documented via Google Maps PDF
- Keep all documentation until you receive your reimbursement!
- Reimbursement Timing
6–8 weeks after all documentation is sent to ICERM. All reimbursements are reviewed by multiple Brown University offices.
- Reimbursement Deadline
Submissions must be received within 30 days of ICERM departure to avoid applicable taxes. No submissions are accepted after six months.