Zero-Shot Multimodal Event Spotting in Soccer Broadcasts
Authors: Sadia Kamal, Mahmudul Hasan, Tony Braskich
Organizations: Department of Computer Science and Engineering, University of Maryland Baltimore County, Maryland, USA · Comcast Applied AI, Comcast, Philadelphia, USA
Automatic soccer event spotting supports highlight generation, video indexing, and sports analytics, but existing methods depend on large annotated datasets and soccer-specific training. We present a zero-shot framework that detects, classifies, and temporally localizes events in full-length soccer broadcasts without task-specific parameter training. Each match is segmented into overlapping clips represented by a compact visual storyboard and the aligned commentary transcript, which a general-purpose multimodal model processes jointly. On three manually annotated matches containing 130 events, the best configuration achieves 72.0% F1@5s and outperforms SoccerNet-trained CALF and NetVLAD++ in a cross-dataset setting, while the storyboard representation reduces inference cost by 72% compared with individual frames. We further analyze the effects of model family, input representation, and prompt optimization, and introduce an evidence-grounded excitement-scoring stage that ranks detected events with natural-language explanations for highlight selection.
Figures & tables
Fig. 1 : Proposed zero-shot multimodal event-spotting framework. Full-match broadcasts are segmented into overlapping clips represented by sparse visual storyboards and aligned commentary. A multimodal LLM predicts event class, timestamp, confidence, and evidence, followed by temporal consolidation and excitement scoring.
Model
Training
Input
P@5s
R@5s
F1@5s
mAP@5s
MAE (s)
Pred./match
Cost/match
CALF
Yes (SoccerNet)
Video
12.91
43.06
19.60
31.95
0.94
147
N/A
NetVLAD++
Yes (SoccerNet-v2)
Video
27.23
65.96
38.49
56.82
0.81
103
N/A
GPT-5.5
No
Storyboard
63.85
82.61
72.02
76.24
0.27
56
$18.01
GPT-5.6 Sol
No
Storyboard
60.93
82.45
70.05
74.77
0.47
58
$18.01
GPT-5.6 Terra
No
Storyboard
59.90
77.98
67.75
74.08
0.35
56
$7.20
GPT-5.6 Luna
No
Storyboard
45.83
77.19
57.32
67.16
0.30
74
$0.72
TABLE I : Average event-detection performance across three matches, with task-specific training and per-match inference cost.
Model
Corner
Goal
Kickoff
Miss
Shot
Yellow Card
GPT-5.5
16/22 (72.7%)
18/21 (85.7%)
6/6 (100.0%)
19/26 (73.1%)
35/41 (85.4%)
14/14 (100.0%)
GPT-5.6 Sol
16/22 (72.7%)
17/21 (81.0%)
6/6 (100.0%)
18/26 (69.2%)
37/41 (90.2%)
14/14 (100.0%)
GPT-5.6 Terra
17/22 (77.3%)
17/21 (81.0%)
6/6 (100.0%)
17/26 (65.4%)
31/41 (75.6%)
14/14 (100.0%)
GPT-5.6 Luna
15/22 (68.2%)
18/21 (85.7%)
6/6 (100.0%)
15/26 (57.7%)
33/41 (80.5%)
12/14 (85.7%)
Qwen3.5-4B
14/22 (63.6%)
19/21 (90.5%)
3/6 (50.0%)
15/26 (57.7%)
14/41 (34.1%)
14/14 (100.0%)
Qwen3-VL-8B
7/22 (31.8%)
11/21 (52.4%)
0/6 (0.0%)
0/26 (0.0%)
0/41 (0.0%)
7/14 (50.0%)
TABLE II : Class-wise event recall across the three soccer matches. Each entry reports correctly detected over annotated events for each class.
Fig. 2 : Average detection recall across three matches. GPT-5.5 and GPT-5.6 Sol achieve the highest recall, while NetVLAD++ and CALF obtain 65.96% and 43.06%, respectively.
Input
P@5s
R@5s
F1@5s
MAE
Pred.
Cost
Transcript only
53.85
46.67
50.00
0.68
52
13.56
15 frames only
59.42
68.33
63.57
0.24
69
64.47
15 frames + transcript
67.14
78.33
72.31
0.91
70
64.69
Storyboard + transcript
65.38
85.00
73.91
0.43
78
18.01
TABLE III : Input representation and modality ablation on a single full-length soccer match using GPT-5.6 with a ±5 -second matching tolerance.
Prompt Strategy
Model
P@5s
R@5s
F1@5s
MAE (s)
Pred./match
Cost ($)
Storyboard
GPT-5.6 Sol
60.93%
82.45%
70.05%
0.47
58
18.01
Storyboard-2
GPT-5.6 Sol
62.24%
71.5%
61.30%
0.14
54
18.01
GEPA
GPT-5.6 Sol
55.41%
71.83%
62.17%
0.23
56
20.00
MIPROv2
GPT-5.6 Sol
53.09%
80.37%
63.89%
0.27
65
20.00
TABLE IV : Comparison of manual and automatically optimized prompting strategies across three soccer matches.
Prompt
Model
Corner
Goal
Kickoff
Miss
Shot
Yellow Card
Storyboard
GPT-5.6 Sol
72.7% (16/22)
81.0% (17/21)
100.0% (6/6)
69.2% (18/26)
90.2% (37/41)
100.0% (14/14)
Storyboard-2
GPT-5.6 Sol
63.6% (14/22)
76.2% (16/21)
66.7% (4/6)
65.4% (17/26)
73.2% (30/41)
85.7% (12/14)
GEPA
GPT-5.6 Sol
72.7% (16/22)
81.0% (17/21)
83.3% (5/6)
65.4% (17/26)
68.3% (28/41)
85.7% (12/14)
MIPROv2
GPT-5.6 Sol
68.2% (15/22)
81.0% (17/21)
100.0% (6/6)
73.1% (19/26)
85.4% (35/41)
100.0% (14/14)
TABLE V : Event-level Recall@5s across event classes for different prompt strategies.
Fig. 3 : Per-match event-level confusion matrices for GPT-5.6 Sol with storyboard and transcript input. Rows denote ground-truth classes and columns predicted classes.
Event
N
GPT Avg.
Avg. Conf.
Flow Avg.
GPT-J Avg.
Corner
21
2.66
0.79
2.67
3.38
Goal
27
7.87
0.92
7.70
8.00
Kickoff
11
2.44
0.86
1.45
2.18
Miss
26
5.03
0.85
4.46
4.81
Shot
72
5.81
0.82
4.69
5.96
Yellow Card
18
2.36
0.87
2.56
3.00
TABLE VI : Excitement-score analysis over 175 post-consolidation event predictions from the six benchmark categories across the three evaluated matches.
Match
N
GPT Avg.
Flow Avg.
Flow ρ
Flow MAE
GPT-Judge Avg.
GPT-Judge ρ
GPT-Judge MAE
Match 1
78
5.05
4.46
0.785
1.28
5.28
0.943
0.57
Match 2
43
5.52
4.53
0.683
1.51
5.67
0.908
0.66
Match 3
54
4.74
4.39
0.581
1.38
4.87
0.880
0.72
Overall
175
5.07
4.46
0.716
1.37
5.25
0.926
0.64
TABLE VII : Agreement of GPT-5.4 Mini excitement scores with two LLM-based judges.
Match
Event
GPT-5.4
Conf.
FlowJudge
GPT-5.6
Explanation
Match 1
Goal
8.61
0.91
9.0
9.0
Ferran Torres scores after a cross, with the ball inside the goal and sustained celebration from the scorer and team. The commentator’s disbelief at two goals in two minutes makes the moment highly exciting.
Match 2
Shot
8.62
0.93
6.0
8.0
The goalkeeper saves a close-range Aguero attempt after a City move into the penalty area. The commentator’s “to seal it for City! What a save” makes the moment highly dramatic.
Match 3
Miss
5.83
0.88
4.0
6.0
Bruno Fernandes spots Ortega off his line and attempts a long-range effort toward an open goal, but it goes wide. The near-open-goal setup adds tension, while the miss and muted commentary limit the excitement.
TABLE VIII : Representative excitement-scoring examples from the three matches.
East China Normal University, Shanghai, China · Migu Video Technology Co., Ltd., Shanghai, China · South China University of Technology, Guangzhou, Guangdong, China