AI / Computer Vision Consultant (Project-Based) – Football Match Video Analysis (Architect AI Soccer Vision Engine)
Budget: $10,000 – $20,000 USD
I’m gearing up to build a production-grade video analysis engine for football and need outside expertise for the next two months. My in-house developers will handle day-to-day coding; what I’m missing is a seasoned Computer Vision / Machine Learning mind (or tight-knit duo) who can define a rock-solid architecture, steer the implementation, and keep us on the right technical track.
Top priority
We must nail player detection and tracking down to individual identification from the very start. Ball tracking, event recognition (pass, shot, dribble, etc.) and performance metrics will follow, but everything rests on reliably following each player throughout a full-match broadcast.
Current state
We haven’t begun formal model development yet, so you have a clean slate to shape data pipelines, model choices, and evaluation strategy.
Tool stack expectations
The codebase will live in Python with PyTorch or TensorFlow at the core. OpenCV, Detectron2 / YOLOv5, Deep SORT, and pose-estimation frameworks such as AlphaPose or MMPose are all on the table—feel free to suggest alternatives.
Key deliverables
• System architecture diagram covering data ingestion, preprocessing, model components, tracking logic, and deployment flow
• Model and algorithm recommendations with pros/cons and reference papers or repos
• Training and evaluation plan, including metrics for individual player ID accuracy and occlusion handling
• Hands-on guidance sessions with our devs (screen-share or pull-request reviews) throughout the build
• Final technical validation report summarising results, remaining gaps, and next steps
This engagement is strictly two months, project-based, and focused on tangible outputs rather than exploratory research. If you’ve shipped multi-object tracking systems to production before and can talk confidently about real-world constraints—occlusions, stadium lighting, broadcast camera motion—let’s dig into the details. Short-listed candidates will receive the full spec and sample footage for review.
Top priority
We must nail player detection and tracking down to individual identification from the very start. Ball tracking, event recognition (pass, shot, dribble, etc.) and performance metrics will follow, but everything rests on reliably following each player throughout a full-match broadcast.
Current state
We haven’t begun formal model development yet, so you have a clean slate to shape data pipelines, model choices, and evaluation strategy.
Tool stack expectations
The codebase will live in Python with PyTorch or TensorFlow at the core. OpenCV, Detectron2 / YOLOv5, Deep SORT, and pose-estimation frameworks such as AlphaPose or MMPose are all on the table—feel free to suggest alternatives.
Key deliverables
• System architecture diagram covering data ingestion, preprocessing, model components, tracking logic, and deployment flow
• Model and algorithm recommendations with pros/cons and reference papers or repos
• Training and evaluation plan, including metrics for individual player ID accuracy and occlusion handling
• Hands-on guidance sessions with our devs (screen-share or pull-request reviews) throughout the build
• Final technical validation report summarising results, remaining gaps, and next steps
This engagement is strictly two months, project-based, and focused on tangible outputs rather than exploratory research. If you’ve shipped multi-object tracking systems to production before and can talk confidently about real-world constraints—occlusions, stadium lighting, broadcast camera motion—let’s dig into the details. Short-listed candidates will receive the full spec and sample footage for review.