Most reported behavioral effects should be treated as preliminary, as the effect size tends to diminish upon replication of the study, sometimes to zero.
I agree that better studies would be based on real life observations by detached observers using video. Even then, doing social science research in a scientific fashion and getting reliable results is hard and might not be forthcoming in my lifetime.
At the same time, I wonder about the relevance of the effect, if it turns out to be real. I don't care if you dehumanize me. I just don't want you to run me over. And I have a hunch that in most crashes, the driver doesn't "see" the cyclist, or they otherwise misjudge the situation with no time to form a conscious personal opinion about me. If I'm invisible, it doesn't matter what I'm wearing.