Voxel51, the leading solution for building visual AI applications, today announced the milestone release of version 1.0 of open source FiftyOne. FiftyOne Open Source provides the foundation for Voxel51’s commercial solution, FiftyOne Teams, together bringing unparalleled ease and efficiency to how visual data is used to develop robust and reliable visual AI applications.
Visual data, which already makes up over 60% of all data traffic, is critical to building innovative AI applications that understand and interact with the real world. However, AI builders find themselves without tools capable of handling the ever-expanding volumes and diverse modalities of visual data and putting it to use in developing their AI models and applications. Existing fragmented and inflexible tools result in complex, unwieldy, and often manual processes that lead to delays and failures in over 80% of AI projects.
“The biggest determinant of success or failure in those projects is not the models they used or the GPUs they used, but the quality of the data sets that they were feeding to them,” said Brian Moore, co-founder and CEO of Voxel51. “As they say, if you feed garbage into a system, you’re going to get garbage out."
Integrated Solution for Visual AI Builders
FiftyOne was developed by Voxel51’s team of computer vision and machine learning experts to reinvent how visual data is used in developing AI models and applications. With this milestone open source release, FiftyOne delivers on the vision of an integrated solution that can extract valuable intelligence from diverse visual data and automate the iterative processes of using that data to build robust AI models. Exhibiting exceptional ease-of-use, flexibility, and extensibility essential to fit how AI builders work, the milestone provides comprehensive capabilities to develop and deliver visual AI applications successfully.
This release includes:
- Support for additional varieties and forms of visual data – from 3D meshes and scenes to point clouds and geometries – throughout exploration, curation, and model development.
- A powerful customizable and extensible framework that enables building interactive data applications, trigger operations, and custom dashboards using Python that aid in fine-tuning models and datasets, while supporting developers in the way they build their visual AI applications.
- Adding native vector search integration with Elasticsearch to expand the ecosystem of tools that can be used seamlessly with FiftyOne. “Vector search allows you to search by semantics, search by natural language even if you don’t have labels or search by visual similarity. So it’s a key workflow for us,” said Moore.
- Open sourcing of the core machine learning techniques for data and models in FiftyOne Brain that help uncover hidden structure and relationships in visual data by algorithmically assessing uniqueness, similarity, representativeness, and more. “It is exactly the problem of edge cases, failure modes, biases, and gaps in your data set that are going to prevent you from easily getting to production-ready AI,” said Moore. “The ability to know where are those edge cases in your data set and then be able to take action… is absolutely what’s required to take the last mile and get one of these models into production.”
“It’s impossible to develop reliable, trustworthy visual AI when you’re struggling to curate and understand millions of data samples and how to use them to build AI models that will succeed in real-world applications,” said Moore. “This milestone release delivers the solid foundation for innovation in visual AI made possible when every AI builder has the ability to bring quality data and analysis to every step of their development process. FiftyOne’s ease of use, extensibility, and transparency are democratizing best practices for visual AI development, and we’re excited to see our thriving community continue to deliver new value that advances the ecosystem.”
“FiftyOne provides us with a centralized place where we can easily understand data to uncover and resolve problems with our annotations and models. It has been so easy to build with FiftyOne given the numerous integrations and flexibility to dig into the data. No wonder it is quickly becoming the standard in the computer vision community,” said Chris Hall, Data Scientist, Vivint Smart Home.
Harnessing the Promise of Open Source AI
The release of open source machine learning algorithms in FiftyOne Brain demonstrates Voxel51’s continuing commitment to open source AI. The AI ecosystem is at a pivotal crossroads, marked by numerous questions and opinions about the true meaning of open source AI — an uncertainty that could shape the future of this rapidly evolving field. Unlike approaches and vendors that rely on closed, black-box, or only partially open offerings, Voxel51 believes that open source and transparency are critical in all aspects of AI, from models to data to systems.
“Open source makes possible the transparency, collaboration, and rigor needed to deliver continuing innovation in AI,” said Moore. “We are excited to continue investing in open source so that the entire AI ecosystem can collaborate to deliver better and more robust innovation that is reliable and trustworthy.”
Tens of thousands of AI builders and their teams already rely on FiftyOne for a wide variety of use cases to deliver more accurate and robust models, improving team productivity by up to 50% and model accuracy by up to 30%. FiftyOne Teams, Voxel51’s commercial offering based on open source FiftyOne, is helping leading enterprises and organizations including LG Electronics, Berkshire Grey, and Precision Planting to make visual AI a reality.
“The future of AI's multimodal... the vast majority of bits that go through routers on the Internet are visual in nature. It's an even larger order of magnitude problem being able to reason about visual data, which is absolutely essential to truly building intelligent systems,” said Moore.
