Phase-aware pose modeling sharpens throw-distance prediction from broadcast video

Researchers have built a machine-learning framework that predicts how far an elite athlete will throw by reading the athlete’s body movements directly from broadcast video, and they report that explicitly labeling the phases of a throw makes those predictions measurably more accurate. The work, published in Sports Engineering, introduces Throwing4, a curated dataset of 500 clips of elite throwing attempts across four disciplines: javelin, discus, hammer, and shot put. Each clip is paired with two-dimensional pose sequences extracted from the footage, frame-level annotations of technique phases, and the official recorded throw distance, giving the models an objective continuous target to learn from.

The central question the team posed was whether treating a throw as one unstructured motion sequence is the best way to estimate its outcome, or whether the natural biomechanical structure of the movement carries information that generic sequence models leave on the table. To test this, the researchers formulated throw-distance prediction from broadcast-derived pose as a regression problem. They compared a baseline approach that models the full sequence of body poses against a phase-conditioned approach that partitions each attempt into four technique stages and fuses the motion representations learned from each stage before making a prediction.

Across multiple graph-based pose encoders, the phase-aware models consistently reduced prediction error relative to the full-sequence baselines. According to the authors, this pattern indicates that phase-aligned modeling yields a more stable representation of throwing motion than treating the entire action as an undifferentiated stream of poses. In practical terms, the improvement suggests that knowing where in the throw a given body configuration occurs, such as during the preparatory stages or the delivery itself, acts as a useful inductive prior that helps the models map movement to distance.

The study was conducted under realistic broadcast conditions, meaning the pose data derive from footage of the kind available to broadcasters rather than from laboratory motion-capture rigs. That choice matters for scalability: if performance models can work from ordinary video, they can in principle be applied to large archives of competition footage without specialized instrumentation. The authors position the dataset and protocol as a foundation for future work on fine-grained technique analysis, interpretable sports analytics, and robust modeling from broadcast video, extending the utility of the approach beyond score prediction alone.

Subject of Research: Phase-aware pose modeling for predicting throw distances of elite javelin, discus, hammer, and shot put athletes from broadcast video

Article Title: Throwing4: phase-aware pose modeling for throw-distance prediction in elite athletics

Article References: Badatya, B. K., Tiwari, K., Amin, J., & Hegde, R. (2026). Throwing4: phase-aware pose modeling for throw-distance prediction in elite athletics. Sports Engineering, 29(2), Article 42. https://doi.org/10.1007/s12283-026-00573-4

Image Credits: AI Generated

DOI: 10.1007/s12283-026-00573-4

Keywords: sports engineering, biomechanics, pose estimation, machine learning, throwing events, javelin, discus, hammer throw, shot put, regression, broadcast video, sports analytics