The focus of our work was to understand each sport that our customer analyzed, memorize team compositions, and tag/label participating athletes seen from the game photographs to their correct names using bounding boxes and image segmentation techniques.
About The Company
Our client is an award-winning sports analytics AI startup that has analyzed 59 Million+ athletes’ bodies and faces across team sports and race events across the globe.
They use a combination of deep learning, scalable cloud infrastructure, and elastic search. Within seconds their AI identifies all legible jersey/bib numbers, player faces, and branded gear information from photos/video captured by photographers in real time.
Sporting events include marathons, and team sports like football, soccer, ice hockey, baseball, lacrosse, and basketball.
The Tech Stack
We worked on the client’s in-house customized data labeling tool to annotate the images. The annotation techniques used were bounding boxes and segmentation. The real-time communication channel while the game was ongoing, was our Slack channel with the customer where we received links to the imagery and business rules for the time-sensitive task at hand.
Helped by 1 Senior Project Manager, the team comprised 2 Data Annotators in the long term and 3 data annotators for occasional gaming season spikes.
Netsmartz’s Data annotators had to go through 4000+ images in bibs cleaning, 1000+ images in bounding boxes, and numerous images for segmentation daily. The challenge was not only to have a near 100% accuracy but also to give them domain knowledge of every sport.
Netsmartz followed the following process to fulfill the requirements of the client:
Mapping & Identifying Appropriate Resources
We took interviews with approx 10 in-house candidates and selected the most appropriate ones who had a natural interest in Sports. They were additionally interviewed on the skills of the English language and previous fast-paced annotation skills. Thereon, Netsmartz was first put to test with other workforce partners by the customer in a two-week assignment. Netsmartz came across as the No. 1 choice for our customer.
Team Engagement & Continuous Training
The team was provided training for a week before they went into the production environment. The learning curve enabled them to generate 95% accuracy over the first two weeks.
Consistent Client Feedback & Action
Initially the Project Manager met the client daily to gather feedback, and thereon weekly. Due to a time zone difference and alignment of shifts, we were able to garner daily client feedback for improvement and implement it in a way that progress was seen in every day of operations.
The primary achievements of this project were:
- We eased resource budget and operational hassle for our customer’s core team by tagging up to 800 images per shift along with bibs cleaning of 2,000-3,000 images per day. This speed was achieved in the initial 3 weeks.
- In addition to their bread & butter market of marathons, our client moves confidently relying on Netsmartz as they take on team sports such as NBA, Major League Soccer, NFL, Indian Premier League Cricket, and Ice Hockey League. All with speed, accuracy, time zone, and cost efficiency, without fretting about having an expensive team of sports enthusiasts who could handle their data annotation.