Evan Juras, Computer Vision Engineer at BDTI, presents the “Tools and Strategies for Quickly Building Effective Image Datasets,” tutorial at the May 2021 Embedded Vision Summit.
A common pain point when using machine learning for computer vision is the need to manually curate and label large quantities of training images. Depending on the application, thousands to millions of images are needed in order to capture the range of visual environments, angles, lighting conditions, and other variations that may be encountered in the field. To make a model accurate enough for real-world applications, significant effort must be invested in creating a training dataset.
Fortunately, there are tools and strategies to help accelerate the task of creating an image dataset. In this presentation, Juras discusses strategies for quickly building a new dataset for training an object detection model and reviews tools and methods for speeding up the process of curating and labeling images.
See here for a PDF of the slides.