Voxae
A vision language model that segments what you ask for, for drones and robots that have to read a scene at runtime
Ask an aerial photograph a question in plain language. It answers with a region rather than a label, and its confidence is drawn as terrain: contours mark equal certainty, and you set the waterline that decides what counts.
- Sheet
- UAVid, held out
- Bridge
- Qwen2-VL 2B to SAM 2.1
- Survey
- 306 annotated samples
Standing the scene up
Survey record
| Question type | Bridge | Baseline | n |
|---|---|---|---|
| Affordance questions | 0.421 | 0.290 | 222 |
| Referring questions | 0.399 | 0.538 | 84 |
| All | 0.414 | 0.371 | 306 |
Mean IoU against the annotated answer. The baseline wins on referring questions, which name a thing you can point at. The bridge wins on affordance questions, which ask what a surface is good for. Both are printed here because the second result only means something beside the first.
How the sheet is drawn
Your browser sends the photograph to a route handler on this site, which passes it to the model and reads the answer back off a stream. The model is never addressed from the browser and no credentials reach it.
What comes back is the confidence surface itself, about 60 KB of greyscale, not a finished picture. Every contour, every tint, and every number that moves when you drag the waterline is computed here from that one response. The answer arrives once and then costs nothing to re-read.
An uploaded photograph has no annotated answer, so it is drawn but not scored. The five listed scenes are held-out evaluation samples and carry theirs.
What can you do with this
One inference. The waterline is the policy, not the model.
Fetching the cached reading