The classes a BDD100K detector can't see

Here's a small thing that turns out to be a big thing.

Our detection taxonomy has 12 classes. Three of them (autorickshaw, animal, and a vehicle fallback for things that fit nowhere else) have no equivalent in BDD100K's label set. A detector trained on BDD100K doesn't get them wrong. It can't output them at all.

How often does that matter? We counted, straight from our detection annotations. Across the whole dataset, 9.42% of detections fall into those three classes.

The share climbs when conditions get worse: 13.91% at night, 15.82% in fog, 11.10% on highways. In clear daytime it drops to 8.57%. So the gap grows exactly where driving gets harder.

Auto-rickshaws carry most of it. We have 185,383 of them, and they appear in 6,775 of our 8,437 clips. That's 80.3%. On a Delhi road, an auto isn't an edge case. It's the scenery.

One caveat, because we'd want it said to us: these counts come from machine labels, not human-reviewed ones. The pattern is clear enough that we trust the direction, and we'll rerun the numbers on reviewed frames as they come in.