How Hudl Achieved 10x Annotation Speed and Cut Review Time by 40% 

Introducing Customer: Hudl

Hudl is the complete sports analytics ecosystem trusted by teams, coaches, and athletes across professional and amateur sports. From recruitment and performance analysis to high-fidelity video tracking, Hudl's Applied Machine Learning (AML) team powers the data intelligence behind some of the world's most demanding sports analytics workflows, spanning soccer, American football, ice hockey, basketball, and beyond.

At the heart of Hudl's product is video and wearables: sensor feeds, live match footage, and uploaded game film that must be annotated with dense, precise tracking data to train and improve the models.

Problem: Manual Workflows and a Disconnected Annotation Stack

Hudl's AML team had built its annotation workflow using third-party platform that served them well for a time. But as the team's ambitions grew, the fit became less comfortable. The vendor was increasingly focused on other product areas, leaving video annotation lower on its roadmap than Hudl needed it to be.

The platform's manual, frame-by-frame workflow also became a limiting factor as volumes scaled. Integration into Hudl’s prelabelling workflows was also key as requirements scaled. 

At the time, annotation and data science operated as sequential rather than collaborative workstreams. Outputs were passed between teams without shared visibility, aligned ontologies, or joint oversight, a model that limited the feedback loops needed to improve data quality iteratively.  It was an arrangement that made sense early on but that the team had begun to outgrow as the two disciplines became more intertwined.

Finally, the platform lacked instant annotation analytics — capabilities the team didn't need at first, but which would prove essential to managing an annotation operation of the size they were growing into.

Solution: An Integrated, Automated Annotation Platform

Hudl selected Encord as its annotation and data management platform for all AML workloads. The decision was driven by Encord's video-native tooling, the ability to integrate Hudl's own ML models as pre-labelling agents, and the platform's flexibility to grow with the team.

Encord now sits as the central annotation and data management layer in Hudl's MLOps pipeline. Raw camera video, live match sources, and other sources land in S3, is ingested into Encord via SDK-based pipelines, and flows through a multi-stage workflow: pre-labelling via Hudl's own models, human correction and completion by the annotation team, automated validation by a Task Agent, and final review before export to model training pipelines.

The platform is used across a wide range of active projects, including soccer and football object tracking from Hudl's current and prototype cameras, broadcast ice hockey and football tracking, American football end zone cameras, and basketball player and ball detection.

Data scientists are now directly involved in the annotation process, such as co-creating ontologies, reviewing outputs, and assessing metrics, rather than waiting for results to be handed over. 

Key Features Driving Value: Task Agents and Pre-labelling Integration

The ability to connect Hudl's own ML inference pipeline directly into Encord's annotation workflow sits at the centre of the team's strategy. A pre-labelling agent feeds model predictions directly into the annotation queue, drastically reducing the work required from human annotators. A Validation Task Agent automatically flags annotation errors before tasks reach the review stage, ensuring analysts focus on quality rather than error-hunting.

Within the annotation team, SAM v2's automated tracking, including backwards tracking, and SAM v3 for detection have been actively used alongside Hudl's custom pipelines, providing additional tooling flexibility.

Track split/merge and interpolation are also essential for correcting identity switches in multi-object tracking, a critical requirement given the density of annotation across player, ball, and referee tracking in live sports footage.

Encord's analytics features have transformed how the AML team manages annotators and tracks project progress. Managers can view team-level throughput and individual performance, enabling coaching and workload optimisation in real time.

Custom metadata schemas are also used to attach compliance metadata to ingested assets, supporting Hudl's internal data governance requirements.

New Use Cases Unlocked: The Homography Editor

One of the most significant developments in Hudl's use of Encord has been the co-development of a custom homography editor. This is a pitch overlay tool that maps annotation points in video space to their corresponding locations on the pitch.

Annotating the homography (the projection between video and pitch space) had been a challenge for Hudl, previously requiring bespoke internal tooling. Working closely with Encord, a dual-view editor was developed, enabling annotators to click both on the pitch diagram and on the video frame simultaneously.

This capability unlocks a whole class of spatial annotation that was simply not possible through any external product and demonstrates the depth of customisation available within the Encord platform.

Results

10x annotation speed improvement on tracking data: By integrating Hudl's own pre-labelling pipeline into Encord, use cases that previously required approximately 10 hours of manual annotation now take around 1 hour.  This step-change in throughput is enabling significantly more training data to be produced, accelerating model development cycles across the AML team.

40% reduction in review time for Ice Hockey tracking: Prior to Encord's Validation Task Agent, review analysts spent considerable time reviewing the jobs manually searching for errors, rejecting tasks, and returning them to annotators. With the validation agent in place, the same process is now 40% faster. Errors are flagged automatically before reaching review, so analysts can focus on overall quality rather than error detection.

Significant speed uplift from interactive tracking tools: Even without pre-labelling, annotators report that Encord's built-in tracking tools provide a substantial speed improvement compared to previous solutions, a testament to the quality of the core annotation experience.

Better alignment between annotation and data science: Encord has brought the annotation and data science teams into a shared workflow. With jointly created ontologies, collaborative review, and real-time analytics, the two teams now work from the same picture — producing clearer requests, smoother delivery, and faster iteration on models.