End-to-end annotation pipelines to provide
training-ready data
De-identification pipeline
Every hour is privacy-cleared before it leaves our servers. Faces are detected and blurred frame by frame, including in mirrors, across multiple people and inverted frames.
Face anonymization · output
Faces
Blurred on every frame, including partial, reflected and upside-down faces.
Multiple people
Bystanders and co-workers are tracked and anonymized alongside the wearer.
Why it matters
Privacy-cleared footage means the same hour sells to more buyers in more jurisdictions.
A short sample from a messy construction site
The same 16 seconds of tiling work, captured by a six-camera unit capture device with four selected angles. 3D hand skeletons are reconstructed per frame and stay consistent across every view.
Left eye
Right eye
Front left
Front right
SIX CAM · 60 FPS21-keypoint hands · 3D
Semantic segmentation
0:00.0/ 0:16.0
0:00.0
scrape floor mortar back into bucket
transport · right-hand · in-contact
0:01.8
scoop mortar from bucket
grasp · both-hands · in-contact
0:03.5
spread mortar on tile back
fasten · both-hands · in-contact
0:10.5
reload trowel
grasp · both-hands · in-contact
0:11.2
finish mortar coverage
adjust · both-hands · in-contact
0:15.0
move tile to doorway base
transport · both-hands · in-contact
On the way
A vision-language model fine-tuned on hazardous work, with experts in the loop
Generic vision-language models fail on hazardous work because the scenes and actions are not in their training data. We are working on a fine-tuned model for labelling hazardous tasks
Hazard footage · source