I agree with the sentiment, but the details will contain fractal devils. The most unusual pedestrians contain the most important learning experiences, but are also the most significant from a privacy perspective.
Edit: or, to put it another way, you don’t want to find out the hard way your machine vision system just mistook a drunk furry thalidomide victim for a wild coyote, but they might not be too thrilled to be on a database of “freaky humans”.
Depends on your legislation. In Germany you would violate the rights of third parties by filming the road with intentions of sharing it with other third parties while not also making the former third parties unrecognizable (black bar)
I doubt that applies to Moderate resolution LIDAR images which provide ~20 or fewer data points about someones face. https://cdn-images-1.medium.com/max/1600/1*vrB0Jrbt4Sjhkw2g_... Might allow you to detect a long beard or very large hat, but don't allow for identification without other evidence.
These questions will get very interesting when self-driving car manufacturers try to enter markets outside the US. It is highly desirable to have local training data, but any large enough effort to collect that data looks indistinguishable from mass surveillance. In Western Continental Europe this is bound to conflict with their expectation of privacy in public spaces. Meanwhile Russia and China might not be so thrilled about the USA having LIDAR scans of their entire country
Edit: or, to put it another way, you don’t want to find out the hard way your machine vision system just mistook a drunk furry thalidomide victim for a wild coyote, but they might not be too thrilled to be on a database of “freaky humans”.