Distilasi Model Deteksi Pelanggaran Dalam Pertandingan Sepak Bola Berbasis Computer Vision dan Mekanisme Attention

Ajiputra, Valentino Reswara (2026) Distilasi Model Deteksi Pelanggaran Dalam Pertandingan Sepak Bola Berbasis Computer Vision dan Mekanisme Attention. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221017-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Keputusan wasit dalam pertandingan sepak bola berperan penting terhadap jalannya pertandingan, namun berbagai studi menunjukkan bahwa keputusan tersebut kerap dipengaruhi oleh faktor eksternal seperti tekanan penonton dan popularitas tim, sehingga berpotensi menimbulkan bias dan ketidakadilan. Untuk mengurangi bias tersebut, dikembangkan sistem deteksi pelanggaran berbasis computer vision yang mengagregasi fitur dari beberapa klip dalam satu kejadian pelanggaran menggunakan mekanisme attention. Sistem sebelumnya telah menunjukkan bahwa deteksi pelanggaran secara otomatis merupakan hal yang realistis untuk dilakukan, namun masih memiliki keterbatasan karena menggunakan backbone berukuran besar dengan kebutuhan komputasi tinggi dan waktu pelatihan yang lama.
Penelitian ini bertujuan untuk meningkatkan efisiensi sistem deteksi pelanggaran melalui penerapan metode knowledge distillation. Proses distilasi dilakukan pada sistem deteksi pelanggaran dengan memanfaatkan SoccerNet MVFouls Dataset serta SN-MVFoul development kit sebagai dasar implementasi. Evaluasi dilakukan untuk membandingkan ukuran model, kecepatan inferensi, serta performa deteksi pelanggaran antara model baseline dan model hasil distilasi. Selain itu, penelitian ini juga mengeksplorasi pengembangan student model khusus per tugas pelanggaran, integrasi Cross-Task Learning Framework, dan eksperimen penggunaan berbagai fungsi loss guna menganalisis dampak masing-masing skenario terhadap performa sistem secara keseluruhan. Digunakan metrik Accuracy, Balanced Accuracy, Precision, dan Recall untuk mengukur performa sistem.

Hasil penelitian menunjukkan bahwa kombinasi backbone X3D-S, konfigurasi hyperparameter distilasi yang optimal (alpha = 0,7 dan temperature = 3,0), Cross-Task Consistency Learning Framework, serta EMD² Loss berhasil menghasilkan student model yang efisien tanpa mengorbankan performa. Dibandingkan dengan teacher model, model hasil distilasi memiliki jumlah parameter 87% lebih sedikit (35.347.356 menjadi 4.635.485 parameter), kecepatan inferensi 86% lebih rendah (1,487 detik menjadi 0,199 detik), serta peningkatan Balanced Accuracy dari 39,598% menjadi 40,207%.
===================================================================================================================================
Decisions made by the referee during a football match play a crucial part in the flow of the game. However, studies show that those decisions are often influenced by external factors, such as pressure from spectators and the team’s popularity, which can potentially introduce bias and unfairness in refereeing. To reduce these biases, computer vision-based foul detection systems have been developed that aggregate features from different clips of the same foul using an attention mechanism. These systems show that automatic foul detection is a realistic idea, but it still has some drawbacks because it uses a complex backbone with high computational cost and long training time.
This research aims to increase the efficiency of the foul detection system by using a method known as knowledge distillation, a method that transfers knowledge from a bigger and more complex teacher model to one or more simpler and lighter student model/s without suffering a serious performance downgrade. The distillation process is applied to the foul detection system by using the SoccerNet MVFouls Dataset and SN-MVFoul Development Kit as the basis of the implementation. Evaluation is conducted to compare the size, inference speed, and foul detection performance of the baseline model with that of the resulting model. This research also explores the development of specialised student models for each foul task, Cross-Task Learning Framework integration, and usage of various loss functions to analyse each scenario’s effects on the system's performance. Evaluation of the performance will be done based on the Accuracy, Balanced Accuracy, Precision, and Recall metrics.
The results demonstrate that the combination of the X3D-S backbone, optimized distillation hyperparameters (alpha = 0.7 and temperature = 3.0), the Cross-Task Consistency Learning Framework, and EMD² Loss successfully produced a lightweight and efficient student model while maintaining competitive performance. Compared to the teacher model, the distilled model reduces the number of parameters by 87% (from 35,347,356 to 4,635,485), decreases inference time by 86% (from 1.487 s to 0.199 s), and achieves a higher Balanced Accuracy, improving from 39.598% to 40.207%.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Cross-Task Learning Framework, Deteksi Pelanggaran, Knowledge Distillation, Mekanisme Attention, Sepak Bola, SoccerNet ============================================================ Attention Mechanism, Cross-Task Consistency Learning Framework, Football , Foul Detection, Knowledge Distillation, SoccerNet
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Valentino Reswara Ajiputra
Date Deposited:	26 Jul 2026 12:37
Last Modified:	26 Jul 2026 12:37
URI:	http://repository.its.ac.id/id/eprint/136655

Actions (login required)

View Item