Small objects on a background with nothing in it.
Sand is the worst background a detector can be given: no texture to anchor on, no edges, and a person on it is a few dozen pixels of contrast. That is the reason this scene is captured, and the reason its run is the lowest-graded of the nine.
- Camera
- Aerial zones over the shoreline, 1280×960
- Frames in the run
- 300
- Labelled instances
- 22,040
- Classes
- 10
- Validator grade
- B · 89.7/100

What this scene is hard for
The reason it is one of the nine, rather than a prettier map that would have proved less.
- A background with no structure. Nothing in the sand tells a model where an object ends, so a box that is two pixels wrong is invisible to the eye and not to a loss function.
- Extreme scale range in one frame, from a parasol to a person lying under it.
- Palm shadow, which occludes people in a shape a detector will happily learn as an object of its own.
- Water, which is specular, moves, and is the one surface in the nine where the renderer and the identity buffer are most likely to disagree.
Frames from the run
Chosen by eye out of the frames published for this capture, with what each one shows. They are the run's own contact-sheet images at the size the site publishes them, not stills re-exported for this page.

The shoreline on the diagonal, with the water shallow enough to read through to the sand 
The whole spit in one frame: palm grove down the middle and both shores either side of it 
Palms throwing long shadows across the loungers, so half the objects sit in shade and half in full sun 
People and parasols along the water's edge, at the size they appear from the capture altitude 
The same shore at dusk, with the sea taking the top half of the frame and the parasols reduced to dots on the sand 
Near-nadir over open sand: a person here is a shadow and a few pixels of colour
Where this map came from
We did not model it. It is a third-party Unreal environment, dressed with a capture contract, and what NameFrame produced is the frames and the labels — no map and no .uasset is redistributed in any archive on this site. No author is named here because we have no agreement with one; if this scene is yours and you would like it credited, write to hello@getnameframe.com.
The data itself is on the capture’s page
This page is about the place. The archives, the validator’s findings, the class list, the checksums and what is wrong with the run are on the Beach capture, which is the one page on this site that means this dataset.