Kingfisher has spent some time in the water at CMU’s University Center swimming pool and Robotics Institute water tank collecting lots of video footage of game pieces. After annotating these videos with labeled bounding boxes, we were ready to start training our YOLOv3 CNN for object detection. Using an Amazon EC2 instance with 4 GPUs to help speed up the process, we trained the network to identify images found on buoys, torpedo targets, chevron tables, and bins. Testing the weights with Kingfisher back in the water, we can identify the targets with almost 100% certainty!
