Now finding things in a military setting might be more difficult as things get camouflaged. However, additional information can be gleaned by using different parts of the spectrum. Both thermal infrared sensors (cheap ones) exist, while near infrared can be used to differentiate vegetation from green painted patterns.
You can also use (monocular) structure from motion to differentiate objects which are hidden through optical camouflage.
All the above can be achieved using cheap camera sensors, but might indeed require considerable computational power to run in flight (as movement makes the process time critical).
Now getting training data for these machine learning models might be difficult. But then you realize that hyper-realistic games such as War Thunder exist, and games such as GTA have been used to train self-driving car models in the past. So even in absence of real scenery there might be options to train models quickly under various circumstances.
So, it is no surprise that you see posts like this from a well funded weapons producer. Much of the tech and know how is just "out there" (see the outback challenge).
I would conclude that we've silently turned a blind eye to the bad side of this drone tech. It all exists, even in civilian form, but for long it was off limits in the offensive war context.