Video & computer vision
Raw video in. Versioned training set out.
An hour of video is tens of thousands of frames. MLtwist cleans the footage, pre-labels it, has people label and review it frame by frame, and delivers it in your format — our team, yours, or both.
Proof
Video programs we've run
100+ tracked objects per frame
Frame-level tracking of people, vehicles, and containers across tens of thousands of frames per video, with QA done on reconstructed video outside a lagging labeling tool.
ReadRetail · delivery drones10 → 7 minutes per frame
Pre-labeling tuned for unstable, wind-blown footage cut per-frame labeling time 30% on a safety-critical people, pets, and objects taxonomy.
ReadMaritime · collectionNationwide ocean footage, to spec
Vetted boat operators and coastal contributors captured video across sea states, weather, and light — validated before full sessions and anonymized.
ReadHow it works
From footage to a training set
Labeling tools mark objects. The work around them — cleaning, splitting, pre-labeling, review, and packaging — is where video projects stall.
Ingest and clean
Pull footage from your bucket, drop corrupt or unusable files, normalize codecs, and split long captures into workable clips.
Pre-label
Run detection and tracking models tuned to your footage so annotators correct tracks instead of drawing every box.
Label frame by frame
Boxes, tilted boxes, polygons, tracks, and events — kept consistent across frames by automated checks.
Review as video
Reviewers check fully labeled, playable video, not single frames, so drift and ID switches get caught.
Deliver a version
Export in the exact format your training code reads, with a Data ID Card, as a version you can diff and roll back.
What we handle
The footage other pipelines choke on
Unstable and high-motion footage
Drone and vessel video where perspective and scale change every frame.
Dense scenes
Hundreds of objects per frame without the labeling tool grinding to a halt.
Radar and non-visible bands
SAR imagery and video, including oriented boxes that follow each object's heading.
Your taxonomy
Safety-critical classes first, with attributes and events defined by your model's needs.
Need footage you don't have yet? We run real-world collection and synthetic generation too.
Send us a clip
Share a representative sample and your label spec, and we'll scope the work with you.
Also available through Carahsoft and Google Cloud Marketplace.