This year’s CVPR Sports Workshop features over 35 accepted papers exploring how computer vision is being applied to sports, from reconstructing 3D poses in yoga and figure skating to estimating expected goals in soccer and tracking pucks in ice hockey.

In this post, I break down the key themes that emerged, the innovative techniques researchers are using, and what this means for the future of sports technology.

3D Pose Estimation & Human Mesh Reconstruction​

Pose estimation remains the backbone of vision-based sports analysis, and this year brought a wave of work pushing beyond 2D to full 3D mesh reconstruction. Several papers tackled this from different angles:

AthletePose3D introduced a benchmark dataset for 3D pose estimation validated against real athletic movement.

Efficient 2D to Full 3D Human Pose Uplifting presented methods to include joint rotations for a more biomechanically consistent model.

This trend reflects the broader goal of bridging the gap between raw visual data and actionable kinematic insights.

Soccer Analytics Domination​

Soccer continues to be a dominant testbed thanks to SoccerNet, with applications ranging from tactical analysis to game event reconstruction:

Action Anticipation from SoccerNet advanced predictive modeling for gameplay.

From Broadcast to Minimap offered an end-to-end reconstruction pipeline of soccer game states. SoccerNet-v3D: Leveraging Sports Broadcast Replays for 3D Scene Understanding demonstrated how replays can be turned into structured 3D scene data.

The SoccerNet benchmark ecosystem continues to drive innovation and offers a powerful foundation for researchers.

Expanding Beyond Mainstream Sports​

This year also featured a growing number of works in niche or underexplored sports:

TT3D: Table Tennis 3D Reconstruction and Ball Spin Analysis in Table Tennis focused on ultra-fast ball motion and spin estimation.

The Way Up: A Dataset for Hold Usage Detection in Sport Climbing targeted climbing, offering a dataset for hold usage detection.

This broadening of scope suggests an increasing appetite for domain-specific datasets and models tailored to diverse movement styles and equipment.

Benchmarks & Datasets​

The workshop showed a clear emphasis on building high-quality, reproducible datasets:

AthletePose3D, Pose-to-Pose: A New Task and Benchmark for Human Pose Transition in Yoga, and The Way Up: A Dataset for Hold Usage Detection in Sport Climbing released structured datasets for 3D pose estimation and activity analysis.

VNL-STES introduced a volleyball-specific benchmark for spatiotemporal event detection.

Creating specialized datasets continues to be one of the most valuable contributions to this space, often catalyzing downstream model development.

Emerging Techniques​

Several novel modeling trends stood out:

Multi-modal approaches, such as in figure skating, combined pose, video, and scores.

Physics-informed learning and synthetic-to-real transfer (e.g., in table tennis ball spin) are improving generalization.

Gaussian Splatting Transformers (GST) and FineCausal causal modeling are expanding the interpretability and realism of motion synthesis and evaluation.

Several teams emphasized zero-shot and domain-adaptive learning for practical deployment.

These point toward a future where systems can generalize across athletes, sports, and recording environments.

How CVPR 2025 Differs from 2024​

Compared to CVPR 2024, this year’s Sports Workshop showed clear signs of a field maturing toward real-world deployment and broader impact. Here are a few notable differences:

1. Shift Toward Real-World Deployment​

CVPR 2024 largely focused on academic benchmarks and proof-of-concept models. In contrast, 2025 introduced more complete pipelines, such as From Broadcast to Minimap and Virtual Banner Replacement, indicating a shift toward deployable tools for coaching, analytics, and media.

2. Increased Industry Collaboration​

2025 saw a marked rise in papers involving industry partners like TrackMan, Tracab, and EVS Broadcast. These collaborations signal growing commercialization interest and tighter integration of AI into real-world sports workflows.

3. Biomechanics-Informed Modeling​

While 2024 emphasized detection and tracking, 2025 brought more work grounded in physics and biomechanics, such as AthletePose3D and Ball Spin Analysis in Table Tennis using synthetic-to-real transfer methods.

4. Explosion of Sport-Specific Datasets​

The field expanded beyond soccer and general human activity. New datasets targeted niche sports like golf (CaddieSet), rock climbing (The Way Up), yoga (Pose-to-Pose), and volleyball (VNL-STES), supporting more specialized training and evaluation.

5. Focus on Broadcast Video Understanding​

More papers tackled broadcast video analytics in 2025. Tools for camera calibration, game reconstruction, and AR overlays reflect a growing emphasis on viewer experience and real-time insights.

6. Growth in Interpretability and Action Quality Assessment​

Unlike 2024, which focused primarily on recognition and tracking, CVPR 2025 included work on interpretability and action quality, like FineCausal and Comprehensive Figure Skating Assessment. These efforts point toward richer, more nuanced feedback systems.

Where the Field is Headed​

If CVPR 2025 is any indication, the next wave of research will:

1. Bring AI to more types of sports beyond the top leagues.
2. Focus on real-time, wearable-compatible, or mobile-compatible systems.
3. Blur the line between coaching, performance tracking, and fan engagement.
4. Move closer to real-world deployment with robust benchmarks and simulation tools.

We're also likely to see tighter integration between vision systems and biomechanics, especially for injury prevention and recovery support.

The CVPR 2025 Sports Workshop showcased a wide range of innovations. From 3D pose reconstruction to real-time event recognition and broadcast augmentation, these advances demonstrate how visual data can be transformed into actionable insights.

As techniques become more robust, multimodal, and generalizable, we’re seeing the potential for AI to influence not just elite performance, but also training, fan engagement, and injury prevention. The increasing availability of sport-specific datasets and benchmarks is accelerating progress across disciplines.

With stronger connections between biomechanics, physics, and vision, the next generation of sports technology will likely deliver real-time, explainable, and widely accessible tools. -->