We published an open dataset of Indian roads: 645,714 dashcam frames from Delhi NCR, with 3.67M tracked 2D detections and 1.29M segmentation masks across 12 classes.
Free on Hugging Face: https://t.co/p9vHbuUmLq
Full leaderboard, charts and sample frames:
https://t.co/Zae2EEawQv
Want your model on it, or a private evaluation on the roads you ship into? Reply or DM us.
Which model should we add next?
How we scored it:
904 frames from Delhi NCR. Two annotators labelled each one without seeing the other's work. Only boxes both drew count (5,207). Boxes only one drew stay out of scoring.
Faces, plates and painted truck numbers blurred.
Our model boxed the bicycle. It missed the woman walking it.
On 5,000 Delhi frames, annotators added 53% of the final boxes and removed 12% of the model's. Median: 17 s per frame.
Next: 100,000 frames, every box checked by a person.
We opened up 645,714 annotated frames of Delhi traffic.
3.67M tracked boxes, 1.29M segmentation masks, 12 classes. Three of them (autorickshaw, animal, vehicle fallback) don't exist in BDD100K.
Free on Hugging Face. Link below.