Source-linked AI summary
BMASH: Ball-Motion-Aware Soccer Header Spotting
Ahmed Endris Hasen, Muhammad Shahzad Khan, Nikolaos Passalis, Jenni Raitoharju
TL;DR
Soccer header spotting is challenging because contacts are brief and difficult to distinguish from similar airborne-ball actions. The paper adapts Video Swin and adds detector-derived ball-motion features through BMASH, evaluating clips and continuous full matches. BMASH improves clip-level average precision and ROC-AUC over Video Swin-32, while full-video event-level F1 is nearly similar.
Problem
Headers are brief, subtle ball–head interactions that are difficult to distinguish from visually similar airborne-ball actions in broadcast footage.
Source: §1 · p. 2Method
BMASH late-fuses full-frame Video Swin action representations with detector-derived ball-motion information from frame-level detections.
Source: §1 · p. 2, §3.1 · p. 4Results
BMASH-32 raised average precision from 0.9423 to 0.9530 and ROC-AUC from 0.9418 to 0.9507 versus Video Swin-32, while full-video event-level F1 was nearly similar.
Source: §6 · p. 8Takeaways & Limitations
Accurate header spotting is an essential first stage toward scalable video-based head-impact monitoring and downstream biomechanical analysis.
Source: §1 · p. 2Full-video spotting was evaluated on only two held-out broadcast matches, and the ball descriptor does not model full trajectories, speed, acceleration, or player-relative geometry.
Source: §6 · p. 8
Abstract
from arXiv · showhide
Recent advances in computer vision have made broadcast sports videos increasingly useful for event analysis, performance assessment, and player-safety applications. In soccer, however, header spotting remains a challenging problem due to the subtle and short-lived nature of header events. This paper focuses on soccer header spotting: identifying moments in broadcast videos where the ball contacts a player's head. We first adapt and evaluate Video Swin as a strong action-recognition baseline for this task, and then introduce BMASH, a ball-motion-aware fusion framework that integrates detector-derived ball features. BMASH combines Video Swin action representations with ball-presence and motion features from frame-level soccer-ball detection, integrating player-action context with ball dynamics to distinguish headers from visually similar events. We evaluate BMASH using game-level splits with separate test matches and rotating validation folds, considering both centered-window classification and continuous full-video spotting. Results show that Video Swin provides a strong baseline for header spotting, while BMASH improves clip-level AP and ROC-AUC over the corresponding Video Swin baseline. In continuous full-video spotting, BMASH achieves a comparable event-level F1-performance with a different precision--recall trade-off.
1 Introduction
Soccer headers are brief, visually subtle ball–head interactions that are difficult to distinguish in broadcast footage. The paper adapts Video Swin and introduces BMASH, which adds detector-derived ball-motion evidence to visual action features, and evaluates both clip classification and full-video spotting.
1 Introduction
- A header’s brief ball–head interaction occupies few frames and little image area, making it a fine-grained recognition problem.Spotting requires reasoning about player motion, context, and ball dynamics rather than relying only on global scene appearance.Source: §1 · p. 2
- Camera motion, blur, occlusion, and low resolution obscure contacts, while similar airborne-ball actions make headers hard to distinguish.Examples include crosses, clearances, long passes, aerial duels, throw-ins, and shots.Source: §1 · p. 2
- Accurate header spotting is an essential first stage toward scalable video-based head-impact monitoring and downstream biomechanical analysis.Existing heading-exposure estimates rely heavily on manual review or wearable sensing with extensive human verification.Source: §1 · p. 2
- BMASH is a lightweight late-fusion framework combining full-frame Video Swin action representations with detector-derived ball-motion information.Ball presence and movement support visual action features; ball information alone is insufficient because non-headers also involve airborne motion.Source: §1 · p. 2
- Evaluation uses game-level splits with unseen test matches for both labeled temporal-window classification and continuous full-video spotting.The paper adapts Video Swin as a baseline and compares it with EfficientNet3D and TSM-ResNet50 under a unified header/non-header protocol.Source: §1 · p. 2
2 Related Work
Prior work addresses broad soccer action spotting, automated header detection, head-impact analysis, video recognition, and ball tracking. This paper’s related-work discussion positions dedicated binary header spotting as a distinct fine-grained task and motivates combining full-frame video representations with ball-motion cues.
2.1 Action Spotting in Soccer
- SoccerNet [10] and SoccerNet-v2 [8] established and expanded benchmarks for temporal event localization, while later methods model temporal context, dense localization, and long-range dependencies [13], [4],, [31], [33].Unified frameworks such as OSL-ActionSpotting [3] standardize implementations of major spotting algorithms.Source: §2.1 · pp. 2–3
- Recent Ball Action Spotting treats headers as one of several interaction classes, whereas this work targets binary header spotting as a standalone fine-grained task.The cited work notes class imbalance and limited header examples in broader multi-class spotting.Source: §2.1 · p. 3
2.2 Soccer Header and Head-Impact Detection
- A prior broadcast pipeline used ball-centered crops and temporal classification, but reported 21.1% precision on full-match tests despite high header-detection sensitivity.Missed or incorrect ball detections can produce crops that omit the header or focus on an irrelevant region.Source: §2.2 · p. 3
- Wearable-sensor studies quantify heading exposure, but sensors generate substantial false positives, making video confirmation important for reliable estimates.Quantitative video analysis is described as a complementary option for scalable monitoring without wearables and extensive manual annotation.Source: §2.2 · p. 3
2.3 Video Action Recognition and Ball Tracking
- Video Swin uses hierarchical shifted-window attention to model spatiotemporal patterns and is presented as a strong backbone for header spotting [25].Earlier TSM [23] introduced temporal reasoning into 2D convolutional networks by shifting feature channels across adjacent frames.Source: §2.3 · p. 3
- Prior ball-tracking work uses temporal information for small, fast-moving objects [14], [28], motivating tests of compact detector-derived ball-motion descriptors alongside video representations.Soccer ball detection is difficult because the ball is small, fast-moving, blurred, and often occluded,, [21].Source: §2.3 · p. 3
3 Method
BMASH combines full-window player-action understanding with detector-derived ball information to spot headers in short broadcast-video windows. Video Swin captures action context, while ball-motion features provide supporting evidence that the fusion classifier can weigh alongside it.
3.1 Overview of our Method
- BMASH combines Video Swin action representations and detector-derived ball-motion features from the same window to predict header versus non-header actions.The branches provide complementary full-frame action understanding and explicit ball information.Source: §3.1 · pp. 3–4
3.2 Video Swin Action Recognition
- Video Swin processes sampled frames together as a spatiotemporal clip, capturing player motion, body pose, nearby interactions, and scene context.Its two-class scores are summarized across sampled temporal clips for fusion.Source: §3.2 · p. 4
3.3 Ball Detection and Motion Encoding
- Frame-level ball detections are summarized into a 10-dimensional descriptor of detection confidence, visibility, box area, center position, and displacement.When no ball is detected in a frame, its detection values are set to zero.Source: §3.3 · p. 4
3.4 Ball-Motion-Aware Fusion
- The fusion MLP receives concatenated 6-dimensional Video Swin scores and 10-dimensional ball-motion features, using ball evidence as support rather than a standalone decision.The classifier can rely more on action representation when ball evidence is weak or missing.Source: §3.4 · p. 5
- Fusion training freezes both feature branches, optimizes only the MLP with fold-specific weighted binary cross-entropy, and selects any tuned threshold on validation.The header-class weight is the ratio of non-header to header samples in the training fold.Source: §3.4 · p. 5
4 Experiments
Experiments use game-level splits across SoccerNet broadcast matches, with centered-window classification and separate continuous-video spotting. The setup compares video-action baselines with ball-motion-aware fusion and evaluates both threshold-independent and thresholded performance.
4.1 Dataset
- Seven SoccerNet matches plus five added matches are split by game, with a fixed two-match test set of 529 windows.Centered header windows require verified visible contact; visually similar nonheaders isolate window recognition, while continuous spotting is evaluated separately.Source: §4.1 · p. 5
4.2 Compared Methods
- At the 0.5 threshold, generic ball-action header scores yield precision = 0.9230, recall = 0.4303, and F1 = 0.5870 on centered clips.The same model obtains approximately 0.56 AP@1 in full-video spotting; prior header detection reports 21.1% precision.Source: §4.2 · pp. 5–6
- The comparisons include generic ball-action spotting and binary header models, with Video Swin serving as the strongest action-recognition baseline and BMASH adding ball-motion features.The evaluated binary baselines are EfficientNet3D and TSM-ResNet50; BMASH uses Video Swin representations with detector-derived features.Source: §4.2 · pp. 5–6
4.3 Experimental Setup
- Ten models were independently trained across rotating validation matches and tested on the same fixed matches, with Video Swin and BMASH using 16- and 32-frame windows.Each fold selected its test checkpoint by validation performance.Source: §4.3 · p. 6
4.4 Evaluation Metrics
- Clip-level evaluation reports AP, ROC-AUC, F1, precision, recall, and balanced accuracy; continuous-video evaluation uses event-level precision, recall, F1, and false positives per match.Full-video detections use 2.0-second windows at 0.5-second stride, smoothing, non-maximum suppression, and a ±2-second matching tolerance.Source: §4.4 · p. 6
5 Results Analysis
BMASH-32 leads the clip-level ranking metrics, while fold-wise analysis reports consistent performance and qualitative header examples. In full-video spotting, BMASH-32 and Video Swin-32 achieve nearly identical F1 with different precision–recall trade-offs.
5.2 Per-Fold and Qualitative Analysis
- Across folds, BMASH-32 appears in the upper-right precision–recall region with a compact distribution, indicating a stronger balance and stable behavior across game-level validation splits.Figure 3 gives a representative header example, marking the detected ball in red and the player in the header action in blue.Source: §5.2 · p. 6, Figure 4 · p. 7, Figure 2 · p. 7
5.1 Overall Performance
- BMASH-32 reaches 0.9530 AP and 0.9507 ROC-AUC overall, the highest values, while matching Video Swin-32's F1-score.BMASH-32 improves AP from 0.8858 to 0.8966 over Video Swin-16 in the 16-frame setting; Video Swin-32 reaches 0.9423 AP and 0.8636 F1-score.Source: §5.1 · p. 6, Table 3 · p. 7
5.3 Full-Video Header Spotting Analysis
- BMASH-32's full-video F1-score is 0.635 versus Video Swin-32's 0.636; BMASH-32 detects more true headers (175 vs. 171) and has highest recall (0.671).BMASH-16 has the highest precision (0.621) and fewest false positives (97), but lower recall (0.609); dense sliding-window inference also accumulates false positives and temporal localization errors.Source: Table 4 · p. 7, §5.3 · p. 7
- All models achieve higher precision on WB than on BN, where models generally produce more false positives.Table 5 reports full-match results separately for the two broadcast test videos.Source: Table 5 · p. 8, §5.3 · p. 8
6 Conclusion
BMASH combines Video Swin action representations with detector-derived ball-motion descriptors. It improves clip-level average precision and ROC-AUC over Video Swin-32, while full-video event-level F1 remains nearly similar; full-video evaluation covers two held-out matches.
6 Conclusion
- BMASH uses lightweight late fusion to combine Video Swin action representations with detector-derived ball-motion descriptors.Ball motion supports Video Swin-based action understanding.Source: §6 · p. 8
- Average precision rose from 0.9423 to 0.9530 and ROC-AUC from 0.9418 to 0.9507 for BMASH-32 versus Video Swin-32.In full-video spotting, the models achieved nearly similar event-level F1-scores, with BMASH-32 detecting more true headers and obtaining higher recall.Source: §6 · p. 8
- Full-video spotting was evaluated on only two held-out broadcast matches, limiting the evidence available to assess generalization.The authors call for larger and more diverse full-match evaluations.Source: §6 · p. 8
- Future work will expand datasets, investigate trajectory-aware and player-relative ball representations, improve continuous-spotting post-processing, and explore wearable-sensor integration.The proposed sensor integration is for downstream head-impact analysis.Source: §6 · p. 8