The rest of this site describes a generator. This section is about what happens when a model is trained on what it produces and then measured against footage it has never seen. The results are published whether or not they flatter the product, because a result you can only trust when it is good is not a result.
How far does synthetic-only training carry a small edge model, measured against real drone footage rather than against itself?
Each write-up names the file and the commit its figures were measured at, and nothing is recomputed, rounded or restated on the way to the web page. Where a measurement is weak — too few frames, an uncontrolled variable, a comparison between two things that differ in more than one way — that is stated beside the number rather than at the bottom of the page.
The same rule already governs the downloadable captures and the validator that grades them: the checks a run failed ship alongside the ones it passed.
The first experiment here did not meet its goal. A detector trained purely on synthetic captures came out stronger on our own held-out data and weaker on real drone footage than the model it was built to replace, and the write-up says so in its title.
That is more useful to a reader deciding whether synthetic data suits their problem than another chart with an arrow pointing up. It is also the only way the next experiment means anything: a track record where every entry succeeded is not a track record.
