Creating Better Datasets for Training More Robust Models in FiftyOne

Nothing hinders the success of computer vision and machine learning systems more than poor-quality data. Gone are the days of focusing only on the model while assuming the data is given and is static; now, machine learning scientists, engineers and data architects spend significant time working with data. And without effective tools for improving datasets, training robust, deployable models can be time-consuming and inefficient. In this talk, we’ll present FiftyOne, an open-source and enterprise software platform that supercharges machine learning workflows by enabling valuable interactive data visualization along with fast model performance analysis. FiftyOne’s flexible data model and integration-minded architecture provides building blocks to enrich most environments and support use cases like improving data quality, managing annotation workflows, building datasets and analyzing model performance. We’ll explore these workflows and illustrate the value of FiftyOne’s innovative data-centric capabilities for improving the model development life cycle, enabling more performant vision systems.