Yes, this was the case. I had done processing separately for a few groups of particles, combined them, and then did a bit more downstream processing. In my case it seems that I couldn’t re-extract all of them at once due to some kind of difference in upstream processing at the level of 2D classification. I re-extracted them separately, but this still didn’t solve the problem.
Then I ran a job that 2D-classified them all together (didn’t even have to bother picking classes as I knew I wanted all of the particles) and the output of this job was enough to allow me to successfully get the 3D ab-initio job to recognize all of the particles.