Updates
What we shipped and learned.
Dataset releases, annotation work and research notes, newest first.
What our annotators changed on 5,000 frames
They added 53% of the final boxes, deleted 12% of the model's, and took a median of 17 seconds a frame.
An annotation tool for people who look at night frames all day
Why our annotation platform is dark, keyboard-first, and shows every annotator their own numbers.
The classes a BDD100K detector can't see
Auto-rickshaws show up in 80% of our clips. A detector trained on BDD100K has no class for them.
From SD card to dataset: how our pipeline works
Dashcam footage goes in one end, BDD100K-format frames with boxes, tracks, masks and GPS come out the other.
Our Delhi road dataset is open on Hugging Face
645,714 annotated dashcam frames from Delhi NCR, with 3.67M tracked detections and 1.29M segmentation masks.
Why we started with Delhi
Most autonomous driving data comes from Western roads. We wanted data from the roads we drive on.