Structure-from-Motion in the Age of Deep Learning
SFM-DL - ECCV'26 workshop
Malmö, Sweden
8 September 2026
Embedded Files
ABOUT
ABOUT
What does it mean to do SfM well today, when learned matching, priors, and representations coexist with minimal solvers, robust estimation, and bundle adjustment?
Structure-from-Motion (SfM) is undergoing a second revolution. After decades dominated by classical multi-view geometry coupled with robust estimation and nonlinear optimization, modern SfM pipelines increasingly incorporate deep learning at critical stages: feature extraction and matching, correspondence filtering, depth/pose initialization, learning-assisted reconstruction, and even as end-to-end pipelines.
However, the core of SfM remains fundamentally geometric. Problems such as relative pose estimation, camera calibration, triangulation, and bundle adjustment are rooted in projective geometry, with many pipeline stages relying on minimal formulations and carefully designed solvers. These classical components offer interpretability, identifiability, and often predictable failure modes, all properties that are difficult to replicate when using learning-based approaches.
Despite striking empirical advances, learning-based SfM methods still face unresolved challenges: limited cross-domain generalization, brittleness under adverse imaging conditions, and insufficient uncertainty modeling. In contrast, classic pipelines can struggle in low-texture scenes, extreme viewpoint changes, dynamic content, repetitive structures, and other challenging real-world conditions.
This workshop focuses on SfM in the age of deep learning not as a replacement of classical methods but as an evolving design space where learning and geometry must be integrated coherently. We aim to bring together researchers from computer vision, photogrammetry, robotics, machine learning, and mathematics to discuss how to build SfM systems that are not only precise and accurate but also robust, generalizable, efficient, interpretable, and reproducible.
Ultimately, the workshop aims to clarify what should be learned, what should remain explicit and geometric, and how to build SfM systems that are accurate, efficient, and trustworthy for real-world deployment.
Topics
Topics
The workshop will cover (but is not limited to) the following topics:
Classical multi-view geometry, theory, and minimal solvers
Minimal problems in multi-view geometry (relative pose, absolute pose, homographies, trifocal constraints, etc.)
Polynomial formulations, Gröbner-basis and resultant-based solvers, solver generation
Degeneracies, critical configurations, observability, and identifiability
Robust estimation built around minimal solvers (RANSAC variants, LO-RANSAC, MAGSAC, etc.)
Geometric verification and consistency in large-scale reconstruction graphs
Learning-enhanced image matching for SfM
Keypoint detection and description, learned local/global descriptors
Learned matchers (including transformer-based matching)
Correspondence filtering, confidence estimation, learned outlier rejection
Multi-image and long-range consistency constraints
Geometry-aware learning and neuro-explicit SfM
Differentiable optimization, unrolled solvers, learned robust estimators
Constraint-aware networks that respect epipolar/cheirality constraints
Hybrid pipelines (minimal solvers + RANSAC + BA) with learned components
Learning to sample / guide hypotheses in robust estimation (learned RANSAC policies)
Neural priors and modern representations for reconstruction
Learned depth/normal priors for initialization
Neural rendering for multi-view consistency
Implicit representations and integration with multi-view stereo (MVS)
Priors for textureless scenes, repeated patterns, or low-light imagery
Robustness, uncertainty, and reliability
Calibration drift, rolling shutter, dynamic objects
Transparent/reflective surfaces, repeated patterns, degeneracies
Uncertainty quantification and failure prediction; confidence-aware SfM
Generalization across domains and sensors
Cross-dataset evaluation, out-of-distribution behavior, domain adaptation
Terrestrial/drone/aerial/satellite, mobile, endoscopy, industrial inspection, cultural heritage
Evaluation and benchmarks
Fair comparisons, reproducibility, open datasets and open tooling
Protocols that separate improvements due to matching vs. geometry vs. optimization
SfM and its “sisters”: Visual Odometry and Visual SLAM
Learning vs. geometry trade-offs in sequential settings
Loop closure, place recognition, long-term localization
INVITED SPEAKERS
INVITED SPEAKERS
TBD name
TBD affiliation
"TBD title"
TBD name
TBD affiliation
"TBD title"
TBD name
TBD affiliation
"TBD title"
ORGANIZERS
ORGANIZERS
Politecnico di Milano (Italy)
ETH Zurich (Switzerland)
University of Udine (Italy)
KTH (Sweden)
Lund University (Sweden)
Microsoft
CVUT (Czech Republic)
PAPER SUBMISSIONS & PROCEEDINGS
PAPER SUBMISSIONS & PROCEEDINGS
end of June, 2026: Paper submission
end of July, 2026: Notification to authors
Aug 15th, 2026: Camera ready for inclusion in ECCV proceedings
All submissions will be handled electronically via the OpenReview conference submission website: (TBD)
All authors must agree to the policies stipulated below.
All papers will be reviewed by at least three reviewers with double-blind peer-review policy.
TECHNICAL PROGRAM
TECHNICAL PROGRAM
TBD
POLICIES
POLICIES
SFM-DL workshop follows the general ECCV 2026 submission policies. ECCV 2026 uses the Springer Nature Code of Conduct for Book Authors as the basis for many of the policies. We therefore urge you to familiarize yourself with these guidelines before preparing and submitting your work to ECCV 2026.
In submitting a manuscript to ECCV, the authors acknowledge that no paper substantially similar in content has been or will be submitted to another conference or workshop during the review period. Please refer to the Submission Policies on the conference web site for additional details on dual submissions and guidelines concerning prior work.
By submitting a paper to ECCV, the authors agree to the review process and understand that papers are processed by CMT, TPMS (Toronto Paper Matching System), as well as OpenReview to match each manuscript to the best possible area chairs and reviewers. Moreover, the authors agree that papers will be checked for plagiarism using iThenticate.
The authors should be aware that each accepted paper is expected to be presented at ECCV in-person by an author (or an authorized delegate).
All accepted papers will be made publicly available by Springer and/or the European Computer Vision Association (ECVA) no earlier than four weeks before the conference. Authors wishing to submit a patent understand that the paper’s official public disclosure is four weeks before the conference or whenever the authors make it publicly available, whichever is first. More information about ECVA is available at https://www.ecva.net/.
Report abuse