Why we started with Delhi

Look at the datasets that shaped self-driving research: Waymo, nuScenes, Cityscapes, BDD100K. Great work, all of it. Almost none of it comes from roads like ours.

Those roads have lanes people use, signals people follow, and a fairly short list of things that move. Ours have auto-rickshaws, e-rickshaws, cycle-rickshaws, tempos, cattle, and two-wheelers carrying three people and a gas cylinder. The lane is more of a suggestion.

India isn't starting from zero. The IDD family from IIIT Hyderabad gave the field real Indian road data, and we lean on it: our first models were fine-tuned on IDD.

But we wanted a lot of footage, from one city, with GPS and time running through it, so you can follow the same road through day, night, rain and fog. That meant capturing it ourselves.

So we put dashcams on Delhi roads and started recording. The rest of these posts are about what we did with that footage, and what we got wrong along the way.