40k param model beats Yolo26n (at least for small objects) (one-ware.com)

A recent advancement in AI object tracking has demonstrated the effectiveness of a custom CNN model using ONE AI, surpassing the widely-used YOLO26n model in detecting small objects, specifically a tennis ball during gameplay. In a controlled experiment, the ONE AI architecture was tailored for tracking, leveraging specific strategies such as motion capture—utilizing inputs from two frames to enhance detection—resulting in significant improvements over YOLO26n, which struggled with low confidence and double predictions due to its lightweight design and lack of Non-Maximum Suppression (NMS). The performance metrics reveal a stark contrast: the custom ONE AI model, with just 40k parameters, achieved 24 frames per second on standard laptop hardware while maintaining a 456 detection accuracy, compared to YOLO26n’s 2 FPS and 379 detections. This shows not only the efficiency of the custom architecture but also its suitability for real-time applications even on edge devices. By keeping the resolution high and employing intelligent data processing techniques, this approach highlights the importance of specialized AI designs tailored to specific tasks, particularly in environments where small and fast-moving objects are prevalent.