Voxel51

Founded 2018

Curate, visualize, and search multimodal datasets to fix computer vision model failures

Pricing model

Demo

You can use Voxel51 to inspect, curate, and manage multimodal data for computer vision and physical AI models. It organizes raw sensor streams, images, LiDAR point clouds, and video into searchable datasets. You can perform visual similarity searches, filter out low-quality samples, and pinpoint where machine learning models fail in edge cases. The software also provides annotation management, automated quality checks on labels, and detailed model evaluation routines across complex 3D scenes, helping engineering teams organize training datasets for autonomous systems and robotics.

What Voxel51 does

Filter datasets by visual similarity to find model edge casesInspect synchronized multimodal sensor data in a unified viewerRun automated checks to identify mislabeled training examplesTag and export curated dataset slices for retraining runsQuery computer vision datasets using plain text promptsTrack dataset versions across model iteration cyclesBenchmark model predictions against ground truth annotationsSynchronized playback across images, 3D scenes, sensor streams, and mapsNatural language and visual similarity search across multimodal datasetsAutomated quality checks to catch dataset annotation errorsFailure analysis to highlight underperforming classes and edge casesModel-assisted labeling and active learning workflowsDataset slicing and filtering to remove low-quality samplesExtensible plugin system to automate repetitive data actions