FiftyOne Visual AI Data Platform

Voxel51 was founded in 2018 out of University of Michigan research by professor Jason Corso and his PhD student Brian Moore; Voxel51, Inc. is based in Ann Arbor, Michigan. It builds FiftyOne, a data platform that helps teams developing computer vision and physical AI find, fix and evaluate the visual data behind their models.

FiftyOne lets teams search images, video, 3D scenes and sensor streams by natural language, visual similarity or metadata, and plays back synchronized multimodal data with annotations and maps. Curation tools remove low-quality samples, labeling adds automated QA, VLM-assisted labeling and active learning, and model evaluation surfaces failure modes. Agentic workflows automate data discovery and annotation QA, and plugins extend the app. FiftyOne Enterprise adds dataset versioning, role-based access and support for billions of samples, deployable in cloud, hybrid, on-premise or air-gapped setups. The core FiftyOne library is open source.

• Multimodal search: natural language, visual similarity and metadata search across samples, episodes and behaviors.
• Visualization: synchronized playback of images, 3D scenes, sensor streams, maps and annotations.
• Curation: filters and embeddings to remove low-quality or redundant samples.
• Labeling and QA: VLM-assisted labeling, automated annotation QA and active learning.
• Model evaluation: analysis of model failure modes and agentic data workflows.
• Enterprise deployment: versioning, RBAC and cloud, hybrid, on-premise or air-gapped options.

EV makers and automotive software centers in Vietnam can curate driving data, find edge cases and run regression tests for perception models. Robotics and factory automation teams can analyze failure cases of pick-and-place or inspection models. AI outsourcing firms can use the open-source library to deliver higher-quality curated datasets to overseas clients.

Customer results on the site include Volkswagen shortening AI regression test cycles by up to 85%, Berkshire Grey speeding up robotic arm investigations 3x, Safely You maintaining 99% fall detection rates and RIOS eliminating repetitive manual transformations on 20 TB+ of visual data. Curating data before labeling avoids paying to annotate samples that do not improve models. Air-gapped and on-premise options suit automotive and defense data rules.

Physical AI for autonomous vehicles, robots and smart cameras generates petabytes of multimodal sensor data, and the bottleneck has moved from model architecture to finding the right data to label and test. Vietnam's automotive, electronics and robotics sectors, including domestic EV makers and global suppliers with local R&D centers, are building perception teams that need such data tooling.

Founded in 2018 at the University of Michigan; Voxel51, Inc. is a Delaware corporation based in Ann Arbor. Raised a USD 30 million Series B in May 2024 led by Bessemer Venture Partners with Drive Capital, Top Harvest Capital, Shasta Ventures and ID Ventures. Official partner for Open Images V7. Listed on the Inc. 5000 in August 2026.