[Paper Review] Sports Re-ID: Improving Re-Identification Of Players In Broadcast Videos Of Team Sports

Bharath Comandur|arXiv (Cornell University)|Jun 6, 2022

Video Analysis and Summarization4 citations

TL;DR

This paper proposes a hierarchical data sampling strategy and a simple Euclidean centroid loss to improve player re-identification in team sports broadcast videos, where players wear similar uniforms, have few training samples, and appear at low resolution. The method boosts mAP by 7–11.5 and R1 by 8.8–14.9 across both CNNs and vision transformers, achieving state-of-the-art performance on the SoccerNet Re-Identification Challenge 2022 leaderboard with 86.0 mAP and 81.5 R1 on the test split.

ABSTRACT

This work focuses on player re-identification in broadcast videos of team sports. Specifically, we focus on identifying the same player in images captured from different camera viewpoints during any given moment of a match. This task differs from traditional applications of person re-id in a few important ways. Firstly, players from the same team wear highly similar clothes, thereby making it harder to tell them apart. Secondly, there are only a few number of samples for each identity, which makes it harder to train a re-id system. Thirdly, the resolutions of the images are often quite low and vary a lot. This combined with heavy occlusions and fast movements of players greatly increase the challenges for re-id. In this paper, we propose a simple but effective hierarchical data sampling procedure and a centroid loss function that, when used together, increase the mean average precision (mAP) by 7 - 11.5 and the rank-1 (R1) by 8.8 - 14.9 without any change in the network or hyper-parameters used. Our data sampling procedure improves the similarity of the training and test distributions, and thereby aids in creating better estimates of the centroids of the embeddings (or feature vectors). Surprisingly, our study shows that in the presence of severely limited data, as is the case for our application, a simple centroid loss function based on euclidean distances significantly outperforms the popular triplet-centroid loss function. We show comparable improvements for both convolutional networks and vision transformers. Our approach is among the top ranked methods in the SoccerNet Re-Identification Challenge 2022 leaderboard (test-split) with a mAP of 86.0 and a R1 of 81.5. On the sequestered challenge split, we achieve an mAP of 84.9 and a R1 of 80.1. Research on re-id for sports-related applications is very limited and our work presents one of the first discussions in the literature on this.

Motivation & Objective

• To address the unique challenges of player re-identification in team sports broadcast videos, where players wear highly similar uniforms and training data is extremely limited.
• To investigate why conventional re-identification methods designed for surveillance fail in this domain due to distribution shifts between training and test data.
• To develop a data sampling strategy that reduces the domain gap between training and test data distributions in low-shot, high-similarity settings.
• To evaluate the effectiveness of a simple Euclidean centroid loss versus more complex triplet-centroid losses in low-data regimes.
• To demonstrate that the proposed method generalizes across both convolutional neural networks and vision transformers in sports re-ID.

Proposed method

• Introduces a hierarchical data sampling procedure that groups training samples by identity and camera viewpoint, ensuring balanced and representative batches during training.
• Employs a centroid loss function based on L2 (Euclidean) distances between feature embeddings and their class-specific centroids, without using triplet mining.
• Combines hierarchical sampling with the Euclidean centroid loss to improve embedding clustering and generalization under low-data conditions.
• Uses the same backbone networks (e.g., OSNet, ViT-B/16) and hyperparameters across all experiments to isolate the effect of the proposed components.
• Applies the method to both convolutional networks and vision transformers, showing consistent performance gains across architectures.
• Validates the method on the SoccerNet Re-Identification Challenge 2022 dataset, using both public and sequestered test splits.

Experimental results

Research questions

• RQ1 Does hierarchical sampling reduce the distribution gap between training and test data in low-shot, high-similarity sports re-ID?
• RQ2 Does a simple Euclidean centroid loss outperform more complex triplet-centroid loss functions when training data per identity is severely limited?
• RQ3 Can the proposed method improve re-identification performance across both CNNs and vision transformers in team sports videos?
• RQ4 Why do convolutional networks and vision transformers exhibit different performance trends when using hierarchical sampling without the centroid loss?
• RQ5 To what extent does the proposed method generalize beyond soccer to other team sports with similar visual challenges?

Key findings

• The proposed hierarchical sampling and Euclidean centroid loss combination increases mAP by 7–11.5 and R1 by 8.8–14.9 without modifying the network or hyperparameters.
• The method achieves a state-of-the-art mAP of 86.0 and R1 of 81.5 on the public test split of the SoccerNet Re-Identification Challenge 2022.
• On the sequestered challenge split, the method achieves 84.9 mAP and 80.1 R1, confirming strong generalization.
• The simple Euclidean centroid loss outperforms triplet-centroid loss in low-data regimes, contrary to expectations from prior SOTA methods.
• Hierarchical sampling alone improves mAP by 5.7–7.3 and R1 by 6.9–9.4 for vision transformers, but slightly degrades performance for some convolutional networks.
• The method demonstrates consistent performance gains across both ResNet50-fc512 and OSNet_x1_0, as well as DeiT and ViT architectures, indicating broad applicability.