Lightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
summary
Compares fine‑tuned YOLOv11 (250‑epoch schedule, early stop at 155) and a traditional image‑processing pipeline against on‑prem VLMs (Qwen 3, InternVL‑3.5 with RAG) and cloud Gemini 3 (GPT‑5 testing mentioned) on a Roboflow ball‑detection set (≈12,000 train, 44 test, 41 val) with manual ground truth; evaluation uses accuracy, precision and recall. YOLO achieves high metrics on many frames but YOLOv11 misses multiple balls in crowded scenes, while Gemini 3 identifies all 11 balls in a crowded image; Qwen reports ~4 and InternVL ~3 detections. VLMs show higher recall but incur heavier VRAM and latency costs and deployment limits.
AI-written summary · about
related talks
27mAnalysis of sports scenes using computer vision models
31mCV at the Dutch Tennis Federation: Utilizing YOLO to create insights for coaches
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
35mFootball video analysis using Deep Learning
35mLarge-Scale Video Intelligence
46mML Use Cases at Think Big Analytics
32mApplying Deep Learning for Object Detection from a Telecom Perspective
42mDeep Learning for Automatic Basketball Video Production