Cornerstone Technologies

We're proud to share the work of Sebastian Vîrtopeanu, one of our innovative Data Scientists. His work in advancing computer vision and real-time analytics is part of his bachelor’s thesis.

It all started as a summer project focused on ping pong tracking, where he incorporated rigorous scientific methods to benchmark and enhance his model. He conducted cross-validation studies comparing his own approach with an established research model, providing insights into model performance under varied lighting conditions and table environments. His first approach integrated fundamental computer vision techniques such as HSV filtering and morphological operations, while switching to pre-trained models like YOLO and SAM with GDINO.

What began as a simple experiment aimed at detecting an orange ping-pong ball with a laptop camera quickly evolved into a sophisticated system. His early tests in real-world settings (from backyard games to challenging indoor environments with artificial lighting) led him to confront several key challenges:

• High-Speed Motion & Lighting Artifacts: Rapid ball movement causes significant distortion, while common 50-60 Hz artificial lighting interferes with video capture, necessitating high frame rates to meet Nyquist’s criteria.

• Background Complexity: Similar-looking objects and reflections demanded more than classical computer vision techniques such as frame differencing, HSV filtering, and morphological operations.

To overcome these hurdles, Sebastian integrated advanced deep learning methods—employing pre-trained models like YOLO and techniques such as SAM with GDINO—to improve detection accuracy for both the ball and the ping-pong table. His system now processes images at an impressive 35–45 FPS with around 1 pixel RMSE accuracy for the ball’s coordinates, making it well-suited for real-time applications.

Beyond detection, Sebastian’s project features several innovative modules:
• Trajectory Reconstruction: By continuously tracking the ball’s position using a frame-by-frame coordinate queue, he developed a robust method for reconstructing its 2D and even 3D trajectory, combining top-down and lateral video perspectives.

• Video Synchronization: Leveraging audio analysis to synchronize multiple video streams, his approach ensures precise cross-view alignment.

• Parallel Processing: Implementing a producer-consumer threading pattern, where one thread handles GPU-driven inference and another manages video output, he significantly boosted performance while addressing Python’s GIL constraints.

The potential applications of this technology are extensive. In sports analytics, such a system can provide coaches and athletes with detailed performance metrics and automated officiating support. In robotics and simulation, accurate real-time tracking is vital for developing responsive control systems and immersive training environments.

Congratulations, Sebastian, on this outstanding achievement!