Our Delhi road dataset is open on Hugging Face

It's out. You can download it today, for free, from Hugging Face.

The dataset holds 645,714 frames pulled from 8,437 dashcam clips around Delhi NCR. Every clip comes with its own GPS track. On top of the frames sit 3.67M 2D bounding boxes with track IDs and 1.29M semantic segmentation masks, across 12 detection classes.

Everything exports in BDD100K format. If your tooling reads BDD100K, it reads this.

Now the honest part. Our models produced these labels, and no person has checked them. About a third of the detections sit below 0.5 confidence. Please treat it as a big pseudo-labelled corpus, good for pretraining and exploring, and don't benchmark a detector against it and call the result accuracy.

Why publish it anyway? Because a large pile of real Delhi traffic, with tracks and GPS, is useful even when the labels are rough. And because we'd rather hear what people need than guess.

If you train perception models on Indian roads, go and download it. Then tell us what broke.