Millions of vehicle photographs. Annotated or raw.

The same labels our own models train against — masks, types and grades, checked by the review flow rather than crowd-sourced once and shipped.

• Pixel masks, not bounding boxes, for vehicles, panels and damages
• Damage typing and a severity grade per detection
• Panel-level part identity, so a label belongs to a named component
• Shot position per frame, so a set can be reassembled in order
• A schema that matches the API response, so nothing needs remapping

Teams training or benchmarking their own vision models who would otherwise spend the first two quarters labelling.

The imagery without labels: real capture conditions, phone cameras, forecourts and driveways, at the volume that self-supervised work needs.

• Original frames, not the rendered or enhanced versions
• Walkaround sets kept whole rather than split into loose images
• Real conditions — weather, glare, dirt, awkward angles, mixed devices
• Capture metadata where it exists, so sets can be grouped
• Plate neutralisation available on request, using the same model the API exposes

Pre-training, self-supervised and augmentation work, and anyone who wants to apply their own labelling standard rather than adopt ours.