Field Converter: Geometry-Initialized Temporal Residual Refinement for World-Grounded Player Pose Estimation from Soccer Broadcasts

Forum topic · 小凯 · 2026-09-11

Summary

This arXiv paper (2609.10498) presents Field Converter, a geometry-initialized temporal residual framework for world-grounded 3D player pose estimation from calibrated soccer broadcasts. Unlike methods that only reconstruct poses relative to a player's own body, Field Converter localizes players in a shared metric world coordinate system. The method initializes each player's root position via ray-ground intersection using camera and pitch geometry, then predicts a temporal residual correction from pose, image, camera, and geometric cues. On match-disjoint evaluation sequences, residual refinement reduces root error from 49 cm with geometry alone to 14 cm with a frame-wise MLP and 10 cm with a temporal convolutional network (TCN), while a Transformer achieves a comparable 11 cm. The resulting world-space MPJPE reaches 13.2 cm. Ablation studies show residual prediction clearly outperforms direct global-root regression, and that temporal context matters more than the specific temporal backbone. Failure analysis identifies airborne motion as the main limitation of the ground-based geometric initialization.

Overview

Research area: Computer Vision Authors: Simon Khan, Laurent Gajny, Jennyfer Lecompte, Sébastien Laporte Published: 2026-09-09 arXiv: 2609.10498

Abstract

Recovering 3D human pose from monocular sports broadcasts remains challenging when players must be localized in a shared metric world coordinate system rather than only reconstructed relative to their own body. The authors introduce Field Converter, a geometry-initialized temporal residual framework for world-grounded 3D player pose estimation from calibrated soccer broadcasts.

Method

1. Geometric initialization: Use camera and pitch geometry to initialize the player root position through ray-ground intersection. 2. Temporal residual refinement: Predict a temporal residual correction from pose, image, camera, and geometric cues.

Results

On match-disjoint evaluation sequences:

Geometry alone: 49 cm root error
Frame-wise MLP: 14 cm
TCN: 10 cm
Transformer: 11 cm (comparable)
World-space MPJPE: 13.2 cm

Findings

Residual prediction clearly outperforms direct global-root regression.
Temporal context matters more than the specific temporal backbone choice.
Failure analysis identifies airborne motion as the main limitation of the ground-based geometric initialization.

---

*Auto-collected on 2026-09-11.*