ExposureEngine: Oriented Logo Detection and Sponsor Visibility Analytics in Sports Broadcasts

ExposureEngine is an end-to-end system that addresses the limitations of traditional horizontal bounding box methods by utilizing oriented bounding boxes and a language-driven agentic layer to provide accurate, rotation-aware sponsor visibility analytics and automated reporting for sports broadcasts.

The Gist

Imagine you are watching a soccer match on TV. Every few seconds, a giant logo for a shoe brand, a beer company, or a car manufacturer flashes on the screen. For the companies paying millions of dollars to be there, the big question is: "How much did we actually get seen?"

Traditionally, answering this question was like trying to count how many times a specific bird flew past a window by hiring a person to sit there with a stopwatch and a notepad. It was slow, boring, and prone to human error.

Then, computers tried to help. But early computer programs were a bit clumsy. They used square boxes (like picture frames) to try and grab the logos. The problem? Soccer logos aren't always straight. They get stretched, tilted, or squashed because the camera is moving, the player is running, or the jersey is wrinkled.

If you try to put a square box around a tilted logo, you end up grabbing a lot of empty space (the grass, the sky, or the player's skin) along with the logo. It's like trying to fit a diagonal piece of toast into a square sandwich cutter; you end up cutting off the crusts and including a lot of the plate. This makes the computer think the logo is bigger and more visible than it really is.

Enter "ExposureEngine"

The authors of this paper built a new system called ExposureEngine. Think of it as a smart, shape-shifting detective that doesn't just look for logos; it understands how they are sitting on the screen.

Here is how it works, broken down into simple parts:

1. The "Rotating Box" Trick (OBB vs. HBB)

Instead of using a rigid, square box, ExposureEngine uses Oriented Bounding Boxes (OBB).

• The Old Way (HBB): Imagine trying to wrap a gift that is lying diagonally on a table using a square piece of paper. You have to fold in huge corners, wasting paper and covering things you didn't mean to.
• The New Way (OBB): Now imagine using a piece of paper that can rotate and stretch to fit the gift perfectly, hugging the edges tightly.
  This allows the system to measure exactly how much of the screen the logo actually takes up, ignoring the background noise.

2. The "Training Camp" (The Dataset)

To teach this detective how to spot logos, the researchers created a massive training camp. They took 1,103 snapshots from real Swedish soccer games. They manually drew those perfect, rotating boxes around 670 different logos (from Adidas to local sponsors). This taught the computer that a logo can be upside down, sideways, or squished, and it still needs to be counted accurately.

3. The "Super Brain" (The AI Model)

They trained a powerful AI (based on a model called YOLOv11) on this data. It learned to spot logos with incredible precision.

• The Result: It's like having a security guard who never blinks. It found 96% of the logos it was looking for, and when it said, "That's a logo," it was right 96% of the time. It's fast enough to watch a live game and count logos in real-time.

4. The "Concierge" (The Language Agent)

This is the coolest part. Usually, you'd have to look at a spreadsheet of numbers to see the results. But ExposureEngine has a Concierge (powered by AI language models).

• You can just talk to it like a human.
• You could ask: "Hey, how much time did the Nike logo get during the second half?" or "Show me a 5-second clip of the Adidas logo right after a goal."
• The system understands your question, digs through the data, and gives you the answer or even creates a video clip for you to share on social media.

Why Does This Matter?

• For the Sponsors: They get a fair bill. They aren't paying for "fake" visibility caused by a computer grabbing too much background. They know exactly how much their brand was seen.
• For the Broadcasters: They can sell ad space more accurately.
• For the Fans: It's a behind-the-scenes look at the business of sports, made easy to understand.

The Bottom Line

ExposureEngine is like upgrading from a clumsy, square-shaped ruler to a flexible, 3D measuring tape. It takes the messy, chaotic world of a live soccer broadcast and turns it into clean, accurate, and easy-to-read data, all while letting you ask questions in plain English. It's not just about counting logos; it's about understanding the shape of the value they provide.

Technical Summary

Here is a detailed technical summary of the paper "ExposureEngine: Oriented Logo Detection and Sponsor Visibility Analytics in Sports Broadcasts."

1. Problem Statement

Quantifying sponsor visibility in sports broadcasts is a critical marketing task traditionally hindered by manual, subjective, and unscalable analysis. While automated computer vision systems exist, they predominantly rely on Horizontal Bounding Boxes (HBB).

• The Limitation of HBB: In dynamic sports footage, logos often appear rotated, skewed, or warped due to camera angles, jersey folds, and perspective distortion. HBBs, being axis-aligned, fail to fit these shapes precisely. They inevitably include significant amounts of background pixels, leading to:
  • Overestimation of logo size and screen coverage.
  • Inaccurate visibility metrics (e.g., exposure duration, prominence).
  • Reduced precision in downstream tasks like logo masking or removal.
• Data Gap: There is a lack of open, annotated datasets specifically for soccer broadcasts that provide Oriented Bounding Box (OBB) annotations, which are necessary to train models capable of handling rotation.

2. Methodology

The authors propose ExposureEngine, an end-to-end system designed for rotation-aware sponsor visibility analytics. The system architecture consists of four main stages:

A. Dataset Construction

• Source: 97 highlight clips from the 2024 Swedish men's elite soccer league.
• Composition: 1,103 curated frames sampled at 1 FPS, covering 32 matches and 16 teams.
• Annotations: 670 unique sponsor logo classes annotated with OBBs (capturing geometry and rotation) using Label Studio. The dataset exhibits a long-tail distribution of logo frequencies.
• Split: 80% training, 10% validation, 10% testing.

B. Detection Model (The Core)

• Architecture: Based on YOLOv11 (specifically the Medium variant), modified to predict Oriented Bounding Boxes (OBB) instead of standard HBBs.
• Loss Function Optimization: To address class imbalance (670 classes with a long-tail distribution), the authors replaced the standard Binary Cross Entropy (BCE) loss with Varifocal Loss (VFL).
  • VFL down-weights easy negative examples to prevent gradient domination.
  • It up-weights positive examples based on localization quality (IoU), aligning class confidence with box accuracy.
• Training: Trained on 3x NVIDIA A100 GPUs for 200 epochs with input resolution of 1280×720.

C. Analytics Pipeline

• Data Processing: Aggregates frame-level detections, applies temporal/spatial filtering, and calculates metrics.
• Metric Definitions:
  • Exposure: Total duration (seconds) a brand is visible.
  • Coverage: Calculated as the ratio of the OBB area (clipped to the frame) to the total frame area.
  • Tightness Ratio (TR): A novel metric comparing OBB area to the enclosing HBB area to measure geometric efficiency.

D. Agentic Reasoning Layer

• Technology: Built on LangGraph and Large Language Models (LLMs).
• Functionality: Translates natural language queries into structured actions.
  • Analysis Agent: Generates insights from detection data.
  • Highlight Agent: Compiles video segments based on criteria (e.g., "Unibet exposure > 50%").
  • Sharing Agent: Publishes clips to social media.
  • Coordinator Agent: Orchestrates multi-step workflows.

3. Key Contributions

1. OBB Logo Detection Model: A state-of-the-art YOLOv11-based detector specifically trained for rotated and skewed logos in sports broadcasts, achieving high precision.
2. Public Dataset: The release of the first open dataset for soccer broadcasts featuring OBB annotations for 670 unique sponsor logos, addressing a significant gap in the research community.
3. Analytics Dashboard: A comprehensive visualization system providing synchronized video players, exposure timelines, and statistical charts.
4. Agentic System: A multi-layer agent framework that enables non-technical users to query complex visibility data and generate reports/media content via natural language.

4. Results and Evaluation

The system was evaluated on the held-out test set and compared against HBB-based baselines.

• Detection Performance:
  • YOLOv11-Medium (OBB) achieved the best results: mAP@0.5 of 0.859, Precision of 0.96, and Recall of 0.87.
  • It outperformed YOLOv8 variants and larger YOLOv11 models in terms of the precision-recall balance.
• OBB vs. HBB:
  • Accuracy: Detection accuracy (mAP) was statistically similar between OBB and HBB models (difference < 1%).
  • Geometric Precision: The critical advantage of OBB is Tightness Ratio (TR). HBBs included excessive background for rotated logos, whereas OBBs provided tight fits.
  • Orientation Necessity: As logo rotation increased (up to ~60°), the efficiency of HBBs dropped significantly (TR ~0.40), while OBBs maintained high geometric alignment.
• Inference Speed:
  • On GPU (G4dn.xlarge), the system achieves ~20 FPS, suitable for near real-time analysis.
  • On CPU (MacBook M3), it runs at ~6.7 FPS.
• Alignment: Over 96.8% of predictions matched ground truth with an IoU > 0.5, and 63.4% achieved very tight alignment (IoU > 0.9).

5. Significance and Impact

• Marketing Accuracy: By eliminating background noise from bounding boxes, ExposureEngine provides auditable and interpretable metrics for Return on Investment (ROI) calculations, preventing the overestimation of sponsor value common in HBB systems.
• Scalability: The system automates a process that was previously manual and subjective, making it scalable for high-volume broadcast analysis.
• Future-Proofing: The integration of LLM agents allows for flexible, user-driven analytics without requiring code changes, bridging the gap between raw computer vision data and business intelligence.
• Research Advancement: The release of the OBB-annotated dataset sets a new standard for training rotation-aware detectors in the sports domain, encouraging further research into geometrically faithful object detection.

In conclusion, ExposureEngine demonstrates that moving from axis-aligned to oriented bounding boxes is essential for accurate sponsor visibility analytics in sports, offering a robust, scalable, and user-friendly solution for the media industry.