A. K. Mackay
Contents

1.1

§1 Selected work

Sponsor exposure tracker

Detects sponsor logos in F1 broadcast footage with YOLO, as a first step towards measuring each brand's screen time. Zero-shot YOLO-World finds no real logos; a YOLO11s fine-tuned on 69 labelled frames reaches mAP50 0.61 on held-out frames.

2026 – present

Ultralytics YOLO11, YOLO-World, PyTorch, OpenCV, Google Colab

Goal

Sponsors pay for screen time, so the end goal is to measure it: how long and how often each brand is on screen during a broadcast. The first step is finding the logos in every frame.

Zero-shot first

YOLO-World detects objects from text prompts such as “Aramco logo”, so it needs no labelled data. On one minute of F1 footage it found no real logos, not even large, clear trackside boards.

Fine-tuning

  • Data. 90 frames from one minute of an F1 broadcast, labelled by hand: 131 boxes across seven brands, plus 51 frames without logos kept as background.
  • Split. The video is cut into 2-second blocks and every fifth block goes to validation, so near-identical neighbouring frames never end up on both sides.
  • Training. A COCO-pretrained YOLO11s at 1280 px, trained for 150 epochs on a free Colab T4 in about 12 minutes. Datasets are versioned and never overwritten, and training resumes from the last epoch after a disconnect.

Results

On 21 held-out frames (35 boxes) the fine-tuned model reaches precision 0.59, recall 0.60 and mAP50 0.61; YOLO-World scores 0 on all three. The validation set is small, so these numbers are indicative. Large trackside boards and the Microsoft logo on the cockpit are found reliably; small, distant logos and car liveries are the main misses. Inference takes 60–80 ms per frame at 1280 px on an M1 Mac.

Next

Label a whole race in CVAT so every brand appears in both training and validation, then track logos across frames with ByteTrack to turn detections into exposure metrics per brand: screen time, number of appearances and share of the screen.