drones are so cheap comparatively to other munitions. And if the idiots are going to be putting that pattern on things, it seems worth it to tune the detection specifically to the pattern. Program half the drones for the pattern and half not. 🤷

Separating dazzle camo from background noise is a monumental task for any current machine vision method. They pretty much all rely on edge detection to understand the shape of objects, and dazzle camo visually introduces a ton of new fake edges while obfuscating the actual edge of the vehicle by making it repeatedly change color if you trace around the edge.

I've not dabbled in CV yet, but it seems like using something like meta's SAM (https://arxiv.org/pdf/2304.02643), and specifically building a dataset for the dazzle pattern (could be completely synthetic: 3D assets of materiel, dazzle texture, FPV downscaling + noise) would get you part of the way there. Now I kinda wanna give it a tickle and see.

Probably! I've had moderate success before doing the synthetic dataset thing (using blender to add objects of interest to pics) to train machine vision models, so I assume it would work for you too. It's a finicky approach that entails a lot of data cleaning though, unless you have a generation method that's really good at not creating impossible geometry. The problem I'd expect (source: instinct/vibes from working in that field for a few years) with your approach would be overfitting to dazzle camo; anything credible would need acceptable false positive and false negative rates for both dazzle and non-dazzle vehicles

Indeed, I imagine the challenge is in the dataset generation. And why I'd want it to be fully synthetic to seamlessly control the masking and texturing. Same asset but dazzle/non-dazzle textures in a variety of orientations, partial cover, differing terrain, various times of day, weather, etc.

I stumbled on CorridorKey (https://github.com/nikopueringer/CorridorKey) from some youtubers which is what inspired the synthetic dataset idea for doing stuff like this.