September 2022 Summaries

Voxel51 has launched a virtual network of 12 Meetups dedicated to computer vision, targeting data scientists, machine learning engineers, and enthusiasts interested in open-source software and real-world applications. The Meetups, which have rapidly attracted over 1,100 members, will feature a mix of technical topics accessible to beginners and seasoned professionals, with an emphasis on academic and practical insights. Meetings will be held virtually every second Thursday of the month until local co-organizers and venues are secured, with each event featuring two to three diverse speakers. The November 10 Meetup will include talks on scaling autonomous vehicles with computer vision and deploying computer vision applications in retail supply chains. Voxel51 also encourages community engagement and offers opportunities for attendees to contribute, including voting on charity donations and participating as speakers or co-organizers. The Meetups are part of Voxel51's broader efforts to support the computer vision community through their FiftyOne toolset, which aids in dataset curation, model evaluation, and speeding up production workflows.

FiftyOne is an open-source toolset designed to enhance the performance of computer vision models by helping data science teams curate high-quality datasets, evaluate models, and visualize data embeddings. This blog post recaps various tips and tricks for using FiftyOne effectively, such as filtering media without specific tags, visualizing embeddings to identify data patterns, exporting labeled images, and grouping multiple slices of image, video, or point cloud samples. It also covers creating views that isolate individual bounding boxes within images to facilitate detailed analysis of overlapping detections. The article encourages engagement with the FiftyOne community on platforms like Slack and GitHub, offering resources for users to deepen their understanding and utilization of the toolset.

Voxel51 has secured $12.5 million in Series A funding from investors including Drive Capital, Top Harvest Capital, and Shasta Ventures, enhancing its mission to advance data-centric machine learning solutions. Since its inception in 2018, Voxel51 has developed both open-source and commercial software, such as the widely adopted FiftyOne project, which enables users to build and refine high-quality datasets and computer vision models. The company has also introduced FiftyOne Teams, a commercial product facilitating secure collaboration among diverse industries. As the demand for visual data applications grows, Voxel51 aims to maintain its position as a leader in organizing and managing visual data, further developing FiftyOne and FiftyOne Teams to support machine learning systems with transparent and high-quality data. The announcement reflects Voxel51's commitment to expanding its team and continuing its journey with the support of its community, partners, and investors.

FiftyOne 0.17, an open-source machine learning toolset developed by Voxel51, introduces several new features designed to enhance computer vision model performance and dataset management. Notable updates include support for grouped datasets that allow for the organization and visualization of multiview data, including 3D samples and geolocation data, using new interactive tools within the FiftyOne App. The release also features a plugin system for custom app behavior, enabling unique visualizations and data operations, along with integration support for training Detectron2 models, a state-of-the-art library from Facebook AI Research. Community contributions have expanded the FiftyOne Model Zoo, adding models such as OpenAI’s CLIP for zero-shot classification. The FiftyOne community continues to grow, with significant engagement on platforms like Slack and GitHub, reflecting the tool's increasing adoption and collaborative development efforts.

FiftyOne is an open-source machine learning toolset designed to enhance computer vision model performance by facilitating high-quality dataset curation, model evaluation, error identification, and embedding visualization, ultimately accelerating production readiness. The blog discusses various community inquiries and solutions, such as checking and filtering valid annotations when loading COCO-styled object detection data, importing CVAT projects into FiftyOne, and selecting multiple samples in the FiftyOne App using shift and ctrl keys. It also addresses viewing frame-level field information using the JSON viewer and handling object-level embeddings with the FiftyOne Brain component. The blog encourages readers to join the FiftyOne Slack community for further assistance and to explore comprehensive documentation and examples available for these features.