Evaluating neural network for certain task Let us assume I have trained an object detection neural network to detect traffic lights such as here: https://youtu.be/P7j6XFmImAg
Now I would like to verify the model will work well enough "in the wild".
I have made a validation set and calculated mAP and IOU values, both for example ~80% and Precision and Recall ~90%.
How can I confirm that these metric values are sufficient to say the model will work well enough?
From my reading, a common benchmark taken is human accuracy. Then they usually compare the neural network's performance to the human's performance and say "ok, it's better than a human, therefore it should be good enough in the wild."
Unfortunately, I don't have such a benchmark.
Are there any other ways to verify neural network accuracy? 
An idea I had is to run a whole bunch of tests on the road and count the "misses" (False Negatives) and then calculate some kind of "FN occurrences per mile" metric for the network. Then maybe you could extrapolate how many miles you would have to drive to miss a light and if this is high enough for you (because you won't drive that much) then I guess you are ok going with the evaluated model.     
 A: You mentioned "human accuracy" as a metric, and that is a reasonable metric if you can show that the new product identifies obstacles better than humans do it is reasonable to say that your product is better than not having your product. However, in this case it will be really hard to find this data.
Another way to determine whether something is "good enough" is using a risk/consequence matrix. In order to do this you need to estimate the probability of a failure, and the consequences of that failure. For example, if you're providing an automated car an indicator of whether there is a red light at 90% accuracy and running a red light will result in a collision 10% of the time then you end up with a 1% chance of an accident at every intersection which costs between \$15K and \$10M in cost to your company (not to mention how horrible it would be to release a product you expect to kill someone). However, this is an extreme example, and if your product has a much lower risk/consequence score then you can show that the expected value of your product is positive and that you aren't expecting to kill anyone.
In other words, identify as many possible failures as possible. Then identify all the possible consequences for those failures. Then identify the probabilities of those consequences happening. Then calculate the expected value of those consequences. 
The same exercise can be used to estimate how much your product is expected to save people by comparing the expected value (cost) without your product and then the expected value (cost) after your product.
