BMASH: Ball-Motion-Aware Soccer Header Spotting

Ahmed Endris Hasen; Muhammad Shahzad Khan; Nikolaos Passalis; Jenni Raitoharju

cs.CVSep 30, 2026

BMASH: Ball-Motion-Aware Soccer Header Spotting

Authors: Ahmed Endris Hasen, Muhammad Shahzad Khan, Nikolaos Passalis, Jenni Raitoharju

Organizations: Faculty of Information Technology, University of Jyväskylä Jyväskylä, Finland · Faculty of Sciences, Aristotle University of Thessaloniki Thessaloniki, Greece

Abstract

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.

Figures & tables

Figure 1. Overview of the proposed BMASH framework. The action-recognition branch and the ball-detection branch are trained separately using task-specific data. Given a sampled video window, the Video Swin branch processes the full stack of $T$ frames and produces action-score features $\mathbf{f}_{v}$ , while the YOLO ball detector is applied frame-wise and its detections are summarized into compact ball-motion features $\mathbf{f}_{b}$ . The two feature vectors are concatenated and passed to a lightweight fusion MLP to predict whether the input window contains a header. BMASH contains a Video Swin branch for action features and a ball-detection branch for ball features. Their outputs are fused to predict header or non-header.

Split	Non-header	Header	Total
Train/validation pool	1067	947	2014
Test set	268	261	529
Total	1335	1208	2543

Table 1. Distribution of the dataset across the experiment.

Fold	Validation game	Train	Validation
0	Brentford–Bristol City	847/957	100/110
1	Hull City–Sheffield Wednesday	822/957	125/110
2	Leeds United–West Bromwich	879/957	68/110
3	Middlesbrough–Preston North End	820/957	127/110
4	Reading–Fulham	893/957	54/110
5	Stoke City–Huddersfield Town	819/957	128/110

Table 2. Rotating game-level validation folds. Counts are given as header/non-header.

Method	AP	F1	Precision	Recall	Balanced Acc.	ROC-AUC
EfficientNet3D	$0.7061{\pm}0.0544$	$0.6606{\pm}0.0429$	$0.5927{\pm}0.0699$	$0.7931{\pm}0.1809$	$0.6094{\pm}0.0517$	$0.7124{\pm}0.0517$
TSM-ResNet50	$0.5889{\pm}0.0642$	$0.6631{\pm}0.0034$	$0.5008{\pm}0.0083$	$\mathbf{0.9820{\pm}0.0206}$	$0.5138{\pm}0.0153$	$0.5913{\pm}0.0457$
Video Swin-16	$0.8858{\pm}0.0250$	$0.7416{\pm}0.0683$	$0.8786{\pm}0.0463$	$0.6525{\pm}0.1040$	$0.7794{\pm}0.0388$	$0.8831{\pm}0.0247$
BMASH-16	$0.8966{\pm}0.0244$	$0.7961{\pm}0.0575$	$0.8096{\pm}0.0691$	$0.8023{\pm}0.1177$	$0.8013{\pm}0.0416$	$0.8990{\pm}0.0223$
Video Swin-32	$0.9423{\pm}0.0131$	$0.8636{\pm}0.0364$	$0.9022{\pm}0.0280$	$0.8333{\pm}0.0749$	$0.8715{\pm}0.0279$	$0.9418{\pm}0.0139$
BMASH-32	$\mathbf{0.9530{\pm}0.0103}$	$\mathbf{0.8639{\pm}0.0383}$	$\mathbf{0.9152{\pm}0.0290}$	$0.8241{\pm}0.0812$	$\mathbf{0.8735{\pm}0.0289}$	$\mathbf{0.9507{\pm}0.0114}$

Table 3. Performance comparison of all evaluated models. Results are reported as mean

\pm

standard deviation across the validation folds. The best result for each metric is shown in bold .

Figure 2. Per-fold F1-score comparison of the proposed BMASH models and the baseline methods. F1 scores across evaluation folds for the BMASH models and baseline methods

Figure 3. Qualitative examples of header spotting. The red box indicates the detected ball, and the blue box highlights the player in the header action. Examples of header events with the detected ball and the player performing the header highlighted.

Figure 4. Precision–recall results across folds. Points show individual folds and stars show the mean for each model. Precision and recall results across evaluation folds for the evaluated models.

Model	GT	Pred	TP	FP	FN	P	R	F1
BMASH-32	261	290	175	115	86	0.603	0.671	0.635
Video Swin-32	261	277	171	106	90	0.617	0.655	0.636
BMASH-16	261	256	159	97	102	0.621	0.609	0.615
Video Swin-16	261	302	175	127	86	0.580	0.671	0.622

Table 4. Full-video header spotting results across the two test videos, using two second temporal matching.

Video	Model	GT	Pred	TP	FP	FN	P	R	F1
BN	BMASH-32	111	133	71	62	40	0.534	0.640	0.582
	BMASH-16	111	137	75	62	36	0.547	0.676	0.605
	Video Swin-32	111	127	70	57	41	0.551	0.631	0.588
	Video Swin-16	111	160	85	75	26	0.531	0.766	0.627
WB	BMASH-32	150	157	104	53	46	0.662	0.693	0.678
	BMASH-16	150	119	84	35	66	0.706	0.560	0.625

Table 5. Full-match header spotting results on the two broadcast test videos, reported separately for each video.

Explore similar work

SoccerNet 2026 Player-Centric Ball Action Spotting: Per-Player Attention with Agreement-Based Ensembling
Jun 23, 2026Faisal Altawijri, Ismail MathkourTemporal Action Localization Linear Attention
TAHAKOM
SoccerLens: Grounded Soccer Video Understanding Beyond Accuracy
May 10, 2026Ismael Elsharkawi, Ahmed Sait, Silvio Giancola +3Weak Visual Grounding Lens
Department of Computer Science and Engineering, The American University in Cairo · Image And Visual Understanding Lab (IVUL), KAUST
SoccerTrack v2: A Full-Pitch Panoramic Video Dataset for Game State Reconstruction and Ball Action Spotting
Aug 3, 2025Atom Scott, Ikuma Uchida, Kento Kuroda +2
Nagoya University∗ · University of Tsukuba

Jun 23, 2026, cs.CVJ/K move · Enter open · S save

SoccerNet 2026 Player-Centric Ball Action Spotting: Per-Player Attention with Agreement-Based Ensembling

Faisal Altawijri, Ismail Mathkour

TAHAKOM

Temporal Action Localization Linear Attention Ensemble

Open paper AskPDF arXiv

We present our submission to the SoccerNet 2026 Player-Centric Ball Action Spotting challenge, which uses a two-stage pipeline: a Track-Aware Action Detector (TAAD) produces per-player action logits from broadcast video, and a Denoising Sequence Transduction (DST) transformer converts game-state features and TAAD logits into structured event sequences. We improve the TAAD with a temporal transformer that adds cross-frame context, alongside several training fixes. For the DST stage, we introduce a two-stage per-player attention mechanism operating on game-state features, and show that a spatial-first attention ordering (cross-player attention before temporal attention) improves validation Macro-F1 by 1.87%. To exploit architectural diversity, we train four model variants and combine them with a Weighted Event Fusion ensemble that applies agreement filtering to suppress single-model false positives while preserving recall, plus a dedicated exception for the rare tackle class. Our final system improves the challenge Macro-F1 from a baseline of 48.6 to 58.94.

Abstract

Explore similar work

SoccerNet 2026 Player-Centric Ball Action Spotting: Per-Player Attention with Agreement-Based Ensembling

SoccerLens: Grounded Soccer Video Understanding Beyond Accuracy

SoccerTrack v2: A Full-Pitch Panoramic Video Dataset for Game State Reconstruction and Ball Action Spotting