💻 computer science

SoccerNet 2026 Challenges Results

This paper presents the results of the sixth annual SoccerNet 2026 Challenges, detailing the evaluation protocols, leaderboards, and top-performing methods for five vision-based football video understanding tasks that attracted 427 teams and 1,129 submissions.

Original authors: Anthony Cioppa, Silvio Giancola, Håkan Ardö, Mohamad Dalal, Jan Held, Jérémie Ochin, Jiayuan Rao, Karen Sanchez, Renaud Vandeghen, Artur Xarles, Olivier Barnich, Albert Clapés, Mathieu Delvaux, Sergio

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anthony Cioppa, Silvio Giancola, Håkan Ardö, Mohamad Dalal, Jan Held, Jérémie Ochin, Jiayuan Rao, Karen Sanchez, Renaud Vandeghen, Artur Xarles, Olivier Barnich, Albert Clapés, Mathieu Delvaux, Sergio Escalera, Bernard Ghanem, Cédric Hons, Antoine Houet, Sotiris Manitsaris, Tom Michel, Pierre Miralles, Thomas B. Moeslund, Mikael Nilsson, Bogdan Stanciulescu, Marc Van Droogenbroeck, Yanfeng Wang, Weidi Xie, Faisal Altawijri, Mohamed Atef, Semen Budennyy, Vasiliy Chelpanov, Puhua Chen, Yixin Chen, Lechao Cheng, Jianling Chu, Ju-Seong Do, Oleg Durygin, Omar Fetouh, Mirco Fuchs, Youssef Ghallab, Falguni Ghosh, Wonjun Heo, Yufeng Hu, Weixuan Huang, Phuong-Linh Huynh-Ha, Matvey Isupov, Yangguang Ji, Siyuan Jiang, Zhenxiang Jiang, Wonyong Jo, Ho-Young Jung, SeongHeon Kang, MinJae Kim, Youngseon Kim, Jakub Komosa, Artem Konshin, Trung-Hoang Le, Jongmin Lee, Lingling Li, Litao Li, Vadim Linkov, Fang Liu, Haoxuan Ma, Shun Makino, Ismail Mathkour, Konstantin Mitin, Mikhail Moiseev, Takumi Nagaya, Yuki Nakamura, Thanh-Khoi Nguyen, Hoang-Phuc Nguyen, Trong-Thuan Nguyen, Christian Orduz, Kwanyong Park, Fabian Perez, Parthsarthi Rawat, SuHyun Rim, Hoover Rueda-Chacón, Atom Scott, Minori Sugimura, Yuyang Sun, Shengeng Tang, Minh-Triet Tran, Ikuma Uchida, Juan Vanegas, Thanh-Nhan Vo, Jiangtao Wang, Yaxiong Wang, Xiaogang Wang, Ruifeng Wang, Rio Watanabe, Jiali Wen, Yongliang Wu, Di Yang, Xu Yang, Zhuo Yang, Xinyu Ye, Yibo Yu, Zihan Zhai, Yu Zhang, Zhenyu Zhao, Zhun Zhong, Yixi Zhou, Xingyu Zhu, Wenbo Zhu, Julian Ziegler

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, global video game tournament where the "players" are not humans, but computer programs (AI) trying to understand soccer matches. This paper is the official report card for the SoccerNet 2026 Challenges, the sixth year of this competition.

The organizers set up five different "levels" or tasks. Each level tests a different skill the AI needs to master to truly understand the beautiful game. Here is a breakdown of what happened, using simple analogies.

The Five Levels of the Tournament

1. The Crystal Ball (Ball Action Anticipation)

The Challenge: The AI watches 30 seconds of a soccer match and has to guess what will happen in the next 5 seconds.
The Analogy: Imagine you are watching a movie, and the screen goes black for 5 seconds. You have to predict exactly what the characters will do when the picture comes back. Did the player kick the ball? Did they pass it?
The Result: 9 teams played. The winner, a team called FAANTRA-WS, got the best score by using a "two-phase warm-up" strategy (like a runner stretching before a race) and combining different models to handle the uncertainty of the future.

2. The Super-Referee (Player-Centric Ball Action Spotting)

The Challenge: The AI must find specific actions (like a pass, a shot, or a tackle), say exactly when they happened, and identify which specific player did it (including their team and jersey number).
The Analogy: Think of a referee who not only blows the whistle but also instantly writes down: "Player #10 on the Red Team passed the ball at exactly 12:04." The tricky part is that players often block each other (occlusion), making it hard to see who is who.
The Result: 6 teams competed. The winner, PAVE, used a "voting system." They had multiple AI models look at the same moment and only count an action if the models agreed on who did it. This helped them spot the rare "tackle" moves that usually confuse computers.

3. The Time-Traveling Photographer (Novel View Synthesis)

The Challenge: The AI is given a bunch of photos of a soccer field from different angles and must create a new photo from a camera angle it has never seen before.
The Analogy: Imagine you have a 360-degree photo of a room. If you close your eyes and imagine standing in the corner, the AI has to "paint" exactly what that corner would look like, including the grass texture and the players, even though no camera was ever there.
The Result: 65 teams entered. The winner, DENSER, solved a problem where standard AI struggled with low-angle shots (like cameras near the ground). They used a "depth guide" to help the AI understand the 3D shape of the field better, creating a much sharper image than the baseline.

4. The GPS Tracker (Spiideo SoccerNet Synloc)

The Challenge: The AI looks at a single, high-resolution photo of half the field taken from a static camera (like a security camera) and must pinpoint the exact real-world location of every player's hips.
The Analogy: Imagine looking at a map of a city from a drone. You see tiny dots representing cars. The AI has to say, "That dot is exactly at 5th Avenue and 42nd Street," even if the dot is very small and far away.
The Result: 88 teams participated. The winner, SELab, used a "smart tiling" method. Instead of trying to see the whole field at once, they zoomed in on specific areas to find the players, then used geometry to map their positions to the real ground. They were incredibly accurate, beating the baseline by a huge margin.

5. The Sports Quiz Master (Visual Question Answering)

The Challenge: The AI is shown a video or image and asked a multiple-choice question about it. The questions can be about the score, the players' names, or complex game situations.
The Analogy: It's like a trivia game show. The host shows a clip and asks, "Who scored the goal in the 88th minute?" or "What color is the goalkeeper's jersey?" The AI has to pick the right answer from four options.
The Result: 76 submissions were made. The winner, vitomeme, didn't just guess; they used a "routing system." They analyzed the question first: if it was about facts, they looked up the data; if it was about the video, they watched closely. They achieved a 98% accuracy rate, which is nearly perfect.

The Big Picture

In total, 427 teams from around the world submitted over 1,100 entries to this competition.

The paper highlights a few key trends that helped the winners succeed:

Higher Resolution: Giving the AI sharper, clearer images (like switching from a blurry phone camera to a 4K TV) helped them see small details.
Teamwork: The best systems often combined the predictions of several different AI models (like a committee of experts voting) rather than relying on just one.
Using the Rules: The winners didn't just look at pixels; they used the "rules" of soccer, like knowing how the camera is angled or how players move, to make smarter guesses.

The paper concludes that while these AI systems are getting very good, they still struggle with the messiest parts of soccer, like when players are hidden behind each other or when the action happens very fast. The goal of these challenges is to keep pushing the technology forward so we can eventually have computers that understand soccer just as well as a human fan does.

Technical Summary: SoccerNet 2026 Challenges Results

1. Problem Definition and Context

The SoccerNet 2026 Challenges represent the sixth annual iteration of an open benchmarking effort designed to advance computer vision research in sports video understanding. While previous editions focused on tasks like action spotting and camera calibration, the 2026 edition expands the scope to five distinct vision-based tasks that address the unique difficulties of football broadcasts: small, visually similar players, frequent occlusions, sparse events in long untrimmed videos, and abrupt camera viewpoint changes.

The five specific challenges are:

Ball Action Anticipation (BAA): Predicting the timing and class of ball-related actions within a 5-second future window based on a preceding 30-second observation window.
Player-Centric Ball Action Spotting (PCBAS): Temporally localizing and classifying ball-related actions while assigning each action to a specific player via team affiliation and jersey number.
Novel View Synthesis (NVS): Rendering photorealistic images from unseen camera poses in multi-view soccer scenes.
Spiideo SoccerNet Synloc (SSS): Localizing athletes in real-world pitch coordinates from a single calibrated static-camera image.
Visual Question Answering (VQA): Answering multiple-choice questions about football broadcasts using text, image, and video inputs.

The initiative aims to provide shared data, unified evaluation protocols, and public baselines to ensure reproducibility and comparability across methods.

2. Methodology and Challenge Structure

2.1 Data and Protocols

Each task utilized specific datasets and evaluation metrics:

BAA: Used the SN-BAA dataset (English Football League). Evaluation adapted mean Average Precision (mAP) for anticipation, averaging scores across temporal tolerances ( $\delta \in \{1, 2, 3, 4, 5, \infty\}$ ).
PCBAS: Based on the FOOTPASS dataset with 8 action classes. Evaluation used macro F1-score at a low-confidence threshold ( $\tau=0.15$ ), requiring exact matches for action class, team, and jersey number within $\pm 12$ frames.
NVS: A synthetic dataset generated via Blender from real broadcast footage, providing 4K images and COLMAP camera parameters. Evaluation prioritized PSNR for ranking, alongside SSIM and LPIPS.
SSS: Utilized static camera images covering half a pitch. The primary metric was mAP-LocSim, which weights detection and localization accuracy based on real-world distance tolerance ( $\tau=1$ m).
VQA: Based on SoccerAgent and SoccerBench, covering 14 categories (text, image, video). Evaluation was strictly answer accuracy on a 500-question challenge split.

2.2 Leading Methodological Approaches

The winning submissions across tasks demonstrated several recurring technical themes:

Anticipation (BAA): The winner, FAANTRA-WS, adapted the official baseline using high-resolution temporal modeling (448p frames) and a two-phase warm-start training recipe. Key innovations included asymmetric logit ensembling of different RegNetY checkpoints and class-weighted losses to handle rare classes. Other teams improved performance by increasing input resolution and using vision-language models (VLMs) to extract tactical context.
Spotting (PCBAS): The winner, PAVE, enhanced the Track-Aware Action Detection (TAAD) and Denoising Sequence Transduction (DST) frameworks. It introduced per-player attention mechanisms and a voting ensemble of four variants. Crucially, it employed agreement filtering to suppress false positives while preserving recall for rare classes like "Tackle."
Novel View Synthesis (NVS): The winner, DENSER, addressed a critical distribution mismatch where ground-level broadcast views were underrepresented in training. It extended EFA-GS (an alias-free Gaussian splatting method) with camera-height-based loss weighting (up to 5 $\times$ weight for ground views), scale-and-shift-invariant depth supervision, and a three-model pixel-average ensemble.
Synloc (SSS): The winner, Boundary-Aware Adaptive Tiling, utilized a top-down framework. It dynamically expanded fixed-grid crops based on coarse detections to ensure full object containment. It adapted RTMPose-X to predict only the pelvis and its ground projection, enforcing physical coupling via a two-keypoint estimator, and used deterministic ray casting to lift 2D projections to metric world coordinates.
VQA: The winner, Frontier VLMs, employed a task-routed elicitation system. Instead of a single prompt, the pipeline routed questions to four paradigms: knowledge grounding, visual prompting, reasoning decomposition, and temporal understanding. It utilized structured metadata (SoccerAgent/SoccerWiki) and visual prompts (numbered player boxes) to guide the Gemini-3.1-Pro reasoning engine.

3. Key Results

Participation was substantial, with 427 teams submitting 1,129 entries across the five tasks. Only teams submitting reviewed technical reports were included in the final leaderboards.

Ball Action Anticipation: The top team (FAANTRA-WS) achieved a mAPavg of 24.08, outperforming the baseline by 7.32 points.
Player-Centric Ball Action Spotting: The top team (FSITAHAKOMPCBAS-1) achieved a Macro F1@0.15 of 58.94, a 12.5-point improvement over the baseline. The "Tackle" class remained the most difficult category due to occlusion and rarity.
Novel View Synthesis: The top team (Sarthi-GameChanger) achieved a PSNR of 29.89, improving the 3DGS baseline by 3.15 dB. However, the best method did not improve LPIPS over the Triangle Splatting baseline, highlighting a trade-off between PSNR and perceptual quality.
Spiideo SoccerNet Synloc: The top team (SELabSoccerSSS-1) achieved a mAP-LocSim of 97.67 and a Frame Accuracy of 81.91%, significantly outperforming the baseline (77.30 mAP-LocSim).
Visual Question Answering: The top team (vitomeme) achieved 98.0% accuracy on the 500-question split, compared to a random baseline of 25.0%.

4. Significance and Contributions

The paper claims three primary contributions:

Documentation: It provides a single reference for the definitions, datasets, and evaluation scores of the five 2026 tasks.
Benchmarking: It presents the challenge leaderboards on held-out splits and summarizes the original contributions of leading submissions, identifying specific design choices (e.g., ensembling, resolution scaling, task routing) that drove performance.
State-of-the-Art Analysis: It synthesizes recurring methodological themes, noting that progress is driven by richer contextual representations, model ensembles, and the explicit use of domain structures (e.g., camera geometry, tactical features).

The authors emphasize that the challenges successfully reduced the risk of non-reproducible evaluation and lowered entry barriers for teams without large in-house datasets. The results indicate that while significant progress has been made, challenges remain in handling occlusions, rare action classes, and the integration of multimodal reasoning with fine-grained perception. The paper concludes by reaffirming the SoccerNet initiative's commitment to maintaining open datasets and standardized protocols to support reproducible research in sports video understanding.