Re-Identifikasi Pemain Dalam Pertandingan Sepak Bola Berbasis Ensemble Transformer Dan Uji Coba Mekanisme Re-Ranking

Akbar, Nouval Royhan (2026) Re-Identifikasi Pemain Dalam Pertandingan Sepak Bola Berbasis Ensemble Transformer Dan Uji Coba Mekanisme Re-Ranking. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221289-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Re-identifikasi pemain sepak bola merupakan topik penting dalam bidang visi komputer yang menghadapi tantangan kemiripan visual antar pemain satu tim, oklusi, motion blur, ukuran crop yang kecil, dan variasi pose. Penelitian ini mengembangkan sistem re-identifikasi pemain pada dataset SoccerNet-v3 ReID berbasis ensemble Transformer dan melakukan uji coba mekanisme re-ranking pada protokol evaluasi per-action. Beberapa backbone dievaluasi, meliputi CLIP ViT-L/14-336, EVA02-L/14, SigLIP-L/16-256, CLIP ViT-B/16, DINOv2, ConvNeXt-L, dan SOLIDER Swin-B, kemudian distance matrix dari backbone terbaik digabungkan menggunakan bobot yang dituning pada validation split. Konfigurasi terbaik menggunakan empat backbone, yaitu EVA02-L/14, CLIP ViT-L/14-336, SigLIP-L/16-256, dan CLIP ViT-B/16, dengan horizontal flip test-time augmentation. Evaluasi pada internal test split SoccerNet-v3 ReID menunjukkan bahwa metode yang diusulkan mencapai 90,956% mAP dan 88,376% Rank-1, melampaui baseline SoccerNet sebesar 59,11% mAP. Hasil eksperimen juga menunjukkan bahwa re-ranking tidak meningkatkan performa akhir secara signifikan pada protokol per-action SoccerNet. Penelitian ini menunjukkan bahwa ensemble Transformer dengan weighted score-level fusion efektif untuk player re-identification, serta bahwa mekanisme re-ranking dan post-processing perlu diuji sesuai karakteristik protokol evaluasi.
==============================================================================================================================
Soccer player re-identification is an important computer vision task that faces challenges such as high visual similarity between players from the same team, occlusion, motion blur, small player crops, and pose variation. This research develops a player re-identification system on the SoccerNet-v3 ReID dataset using an ensemble of visual foundation models with weighted score-level fusion. Several backbones are evaluated, including CLIP ViT-L/14-336, EVA02-L/14, SigLIP-L/16-256, CLIP ViT-B/16, DINOv2, ConvNeXt-L, and SOLIDER Swin-B, and the distance matrices from the best-performing backbones are combined using validation-tuned weights. The best configuration uses four backbones, namely EVA02-L/14, CLIP ViT-L/14-336, SigLIP-L/16-256, and CLIP ViT-B/16, with horizontal flip test-time augmentation. Evaluation on the internal test split of SoccerNet-v3 ReID shows that the proposed method achieves 90.956% mAP and 88.376% Rank-1, substantially outperforming the SoccerNet baseline of approximately 59.11% mAP. The experiments also show that re-ranking mechanism doesnt give significant improvement for the final performance under the SoccerNet per-action protocol. This research demonstrates that weighted score-level fusion of visual foundation models is effective for player re-identification, and that backbone selection and post-processing methods must be aligned with the characteristics of the evaluation protocol.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Re-identifikasi pemain, SoccerNet, CLIP, EVA02, visual foundation model, weighted score-level fusion, Player re-identification, SoccerNet, CLIP, EVA02, visual foundation model, weighted score-level fusion
Subjects:	T Technology > T Technology (General)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Nouval Royhan Akbar
Date Deposited:	28 Jul 2026 02:10
Last Modified:	28 Jul 2026 02:10
URI:	http://repository.its.ac.id/id/eprint/137978

Actions (login required)

View Item