# Voxel51 A multimodal data platform for finding model failures and improving physical AI training data. - Findings24: - Site: - Preços: Freemium - Categorias: Limpeza de dados com IA, Visualização de dados com IA, Hubs de modelos de IA ## O que é Voxel51? Voxel51 is a multimodal data platform for teams developing visual and physical AI. Built on FiftyOne, it brings dataset search, visualization, curation, annotation, and model evaluation into a connected workflow. Teams can inspect video, LiDAR, images, 3D data, and sensor streams to locate failure cases and refine the data used for training and testing. The platform is available as a commercial offering, alongside an open-source FiftyOne option. ## O que Voxel51 pode fazer? ### Search across multimodal datasets Find samples using natural-language queries, visual similarity, and metadata filters across datasets, episodes, and behaviors. ### Inspect synchronized sensor context Explore multimodal episodes with synchronized playback across images, 3D scenes, sensor streams, maps, and annotations. ### Curate datasets around model behavior Slice, query, and filter large datasets to identify low-quality samples, outliers, anomalies, and data relevant to model failures. ### Annotate with model assistance and review Create labels for images, video, and 3D data; use agentic auto-labeling, active learning, configurable review workflows, and automated quality checks. ### Evaluate models at sample level Compare model predictions with ground truth, inspect failure patterns, and analyze metrics including precision, recall, accuracy, and F1. ### Deploy with governance controls Run in cloud, private, hybrid, or on-premise environments, with dataset versioning, role-based access controls, audit logs, and deployment options that include air-gapped environments. ## Casos de uso ### Investigate physical AI failures Locate underperforming classes and specific failure patterns in video, LiDAR, and sensor data, then examine the surrounding multimodal context. ### Select data for annotation Use model-assisted filters, similarity search, and active learning to prioritize samples likely to improve a model instead of labeling an entire collection indiscriminately. ### Run multimodal labeling projects Coordinate labeling, review, and approval for image, video, and 3D annotations, including synchronized 2D and 3D data. ### Prepare and validate training datasets Find outliers and quality issues, refine dataset slices, and use versioning and model evaluation to assess changes across iterations. ## Perguntas frequentes ### What data modalities does Voxel51 support? The platform lists images, video, 3D point clouds and meshes, LiDAR, geospatial data, audio, DICOM, NIfTI, MCAP, and custom media types. Multimodal views can synchronize 2D and 3D data. ### Can teams use automated annotation? Yes. Annotation workflows include VLM-assisted agentic labeling, zero-shot prediction, active learning, and automated quality checks, with review by people and configurable project stages. ### How is the platform priced? Plans are priced by seats and Voxel Processing Units (VPUs), not data volume. Team, Growth, and Custom plans are listed; specific plan limits vary, and Custom pricing is provided through sales. ### Can Voxel51 be deployed on-premises? On-premises deployment is listed for Growth and Custom plans. Air-gapped deployment is also listed as an option for those plans. ### Does Voxel51 include model evaluation? Yes. Listed capabilities include model comparison, scenario evaluation, aggregate metrics, confusion matrices, and sample-level comparison of predictions with ground truth. ## Informações rápidas - **Product category**: Multimodal data and visual AI platform - **Underlying technology**: Built on FiftyOne - **Primary workflow**: Search, visualize, curate, annotate, and evaluate datasets - **Data types**: Images, video, LiDAR, 3D, sensor, geospatial, and other listed modalities - **Commercial plans**: Team, Growth, and Custom; priced by seats and VPUs