1We helped establish an iterative model training loop: annotate a focused sample, train an initial baseline model, use it to pre-annotate additional data, then selectively add harder and more informative examples. This improved annotation efficiency while also making it easier to target the scenarios that mattered most for production performance.
The workflow also evolved from using Roboflow platform for everything toward a more scalable tooling stack using CVAT for annotation and FiftyOne for dataset management, filtering, and export control.
2Foocus required multiple specialized models for different brands, sports, and customer use cases. profiq worked across dataset preparation, augmentation, training, evaluation, and iterative model improvement.
A key practical takeaway was that augmentation helped, but real production data remained essential. Synthetic or augmented samples improved results only when combined with a strong base of real labeled imagery.
3The engagement also included evaluating which emerging AI techniques could create real value in production. Segmentation-based logo placement analysis distinguished whether a detected logo appeared on a person, such as a jersey, or in the surrounding scene, such as signage. A general logo detector based on SAM3 plus a specific classifier pipeline created a faster path for onboarding new customers with fewer labeled examples. Additional experiments included synthetic data generation in Blender, novel visual backbones (DINOv2 and DINOv3), and vision-language models (Qwen2.5-VL).
4In addition to research and model training, profiq took care of running production inference workflows and helped consolidate training-related scripts and processes, making the overall system easier to maintain and improve over time. On-demand basis, we also provided help with software development, extending the customer facing app.
