Machine Vision with YOLO Object Detection

This is a short demo showcasing the real-time performance of a You-Only-Look-Once (YOLO) object detection model trained on my own custom dataset of entirely white chess pieces.
Proper tuning of the model's hyperparameters enables the model to detect black chess pieces with confidence scores exceeding 60% in many scenarios despite having never seen them before.
The YOLOv7 model was trained in 20 minutes on a GTX 1070 using its 1920 CUDA cores. Inference time was less than 20 milliseconds for a frame rate of about 50 FPS running on a GTX 1070's CUDA cores. The performance is quite extraordinary for old, consumer hardware! For comparison, the NVIDIA A100, an aging commercial GPU, has almost 7,000 CUDA cores.
Future compute and software advancements will certainly take this technology to increasingly exciting places. Perhaps one day soon, AI-driven object detection will be ubiquitous in all kinds of industry from manufacturing and logistics to healthcare and environmental monitoring, transforming how we interact with and manage physical and digital spaces.