Football3D

3D Reconstruction of American Football Game Situations from Handheld Monocular Video

Kanta Sawafuji1 · Rintaro Otsubo1 · Dan Mikami2 · Hideo Saito1

1Keio University   2Kogakuin University

CVPR 2026 Workshop on Computer Vision in Sports

Paper Code Coming soon Video BibTeX
Overview of the football game reconstruction pipeline

From video to 3D. Input handheld video → field registration → player localization → 3D visualization.

01

Abstract

We present a method for reconstructing American football game situations from handheld monocular videos. Unlike professional sports environments that depend on fixed multi-camera systems or instrumented players, our approach targets amateur-level footage captured by a single moving camera. The pipeline estimates field geometry, tracks players, and projects their positions onto the playing surface, then stabilizes the recovered trajectories with pose-aware motion constraints. These estimates support coherent 3D visualization and player-view analysis using only ordinary game video, making spatial game analysis more accessible in settings without specialized capture infrastructure.

02

Method

01
Topology-constrained field registration

Topology-Constrained
Field Registration

Field markings and their known topology anchor each moving-camera frame to a consistent field coordinate system.

02
Player tracking and field localization

Player Tracking and
Field Localization

Detected players are associated over time and projected from image coordinates onto the registered playing surface.

03
Pose-aware trajectory smoothing

Pose-aware
Trajectory Smoothing

Body pose and motion cues suppress projection noise while preserving rapid, sport-specific changes in direction.

03

Results

3D reconstruction. Recovered player locations in a shared world coordinate system.
04

Citation

If you find this project useful, please cite our work.

@inproceedings{sawafuji2026football3d,
  title={3D Reconstruction of American Football Game Situations from Handheld Monocular Video},
  author={Sawafuji, Kanta and Otsubo, Rintaro and Mikami, Dan and Saito, Hideo},
  booktitle={CVPR Workshop on Computer Vision in Sports},
  year={2026}
}