Overview
An international sports data company planned to acquire a technology business that analyzes sports video streams and produces data and insights with computer vision and machine learning. The client asked us to evaluate the target's original plans, its current and future product strategy, the capability and speed of its development process, and its potential to expand into other sports.
Solutions
We joined the client's technical due diligence in its second stage, when the deal team needed an independent technical view of the target. Over eight weeks a nine-person team covering architecture, data, AI, delivery process and quality assurance worked through the target's technology.
- Three days of presentations and Q&A with the target's senior engineers and management team.
- A review of the detection and tracking algorithms and of the code behind them.
- A comparison of the target's computer vision, AI and ML approach with what we see on similar projects.
- A final report with the findings, the risks and the open questions for the deal team.
Features
- Detailed evaluation
- Risk identification
- Algorithm assessment
- Technical interviews
- Product strategy analysis
- Cloud infrastructure review

Business Value
The client received an independent technical picture of the target before deciding on the investment. The report set out the target's strengths in research, development and engineering, and the gaps in product readiness and roadmap planning. The client used it as one input to its decision on the acquisition.
Final Insights
We evaluated the algorithms and the engineering work in detail. The target is strong in computer vision, AI and ML research and development. It has a well-trained technical team, player monitoring and tracking that performed well in our review, and a cloud infrastructure that can scale. Its product, however, is not yet market-ready, and it lacks a clear product roadmap that shows what is ready and what is still in development.
From a technology point of view, we assessed the target as a good fit for the client.
How We Assess Computer Vision Products
A video analysis company sells data, so the assessment starts with the data it produces. We review computer vision products along these lines:
- Detection and tracking quality. We run the target's models on footage they have not seen, preferably supplied by the buyer, and measure detection accuracy, identity switches and track continuity with standard metrics such as HOTA. Results are split by camera angle, lighting, crowding and sport.
- Event detection. Event outputs are compared with a manually labelled sample to measure precision, recall and timing error.
- Data and labelling assets. Size and coverage of the labelled dataset, who owns it, how labels were produced, whether written guidelines exist and how well annotators agree.
- Cost per processed match. GPU time, storage and human review per match, today and at the buyer's projected volume. Manual correction hidden behind an automated product changes the margin.
- Transfer to new sports. What can be reused from one sport to the next and what needs new data, new models and new labelling.
- Latency and infrastructure. Whether the pipeline meets the delivery time customers pay for, from live feeds to post-match reports.
Each finding gets a severity rating and an estimated cost to fix, so the deal team can price it.
Where Sports Video Analysis Still Struggles
Public benchmarks show how far sports video understanding is from solved. The SoccerNet Game State Reconstruction benchmark, published at the CVPR 2024 workshops, asks a system to place every player on a pitch minimap and identify their role, team and jersey number from broadcast video. The authors' end-to-end baseline reached 22.26% on their GS-HOTA metric on the benchmark's test set (on the separate 2024 challenge set, the baseline scored 23.36). They report that it struggled most with jersey number recognition, followed by pitch localization.
For a buyer, this has three consequences:
- Demos are not evidence. A good highlight reel says little about performance in poor broadcast conditions. We test on footage the buyer chooses.
- Identity is harder than detection. Finding players in a frame is the easier part. Knowing which player is which through occlusions and camera cuts is harder, so products that promise player-level statistics need the closest look.
- Human correction is part of the cost. Many sports data products reach commercial accuracy with people reviewing model output. That is a sound approach, but it has to appear in the unit economics.
Our technical due diligence services and computer vision development pages describe the full scope of this work.
from kickoff to final report
of working sessions with the target company’s senior engineers and management
experts on the assessment team