Check-out Recognition This work is done during my internship at Sunmi AI Lab. The primary development environment is using Docker container on a Ubuntu 16.04 machine. The framework used is Pytorch and Tensorflow. Workflow using Jira. Documentation using Confluence Developed a video streaming and preprocessing program by implementing a video parser using OpenCV. Developed image classification algorithm by applying Resnet on video frames. Developed object detection algorithm by applying RefineDet, created a two-stage model by combining Resnet50 with RefineDet512. Improved accuracy by 10.5% by implementing label smoothing, focal loss, learning rate warmup + cosine schedule, Ranger optimizer. Improved training speed by 500%, reduced model size by 33% by implementing distributed data parallel and mixed precision training on Nvidia Titan RTX P2 GPU compared with single GPU, data parallel, full precision setting. Visualized and explained how the model works by applying Grad-CAM, Guided-Grad-CAM.