@nttstar
You have mentioned here that combining ms1m and vggface2 may hurt the performance. How about combining ms1m and glint360k and megaface? Is it a good idea? Of course we should use an automatic process for combination. The idea is that we have a base dataset (for example glint360k) and want to combine it with another dataset (for example ms1m). For each person in ms1m we compare its center embedding with all 360k and if it does not match with anyone (using a low similarity threshold) we can add that person to the base dataset. You may say that all ms1m identities are included in glint360k, but how about megaface? It has 672k identities.
@nttstar You have mentioned here that combining ms1m and vggface2 may hurt the performance. How about combining ms1m and glint360k and megaface? Is it a good idea? Of course we should use an automatic process for combination. The idea is that we have a base dataset (for example glint360k) and want to combine it with another dataset (for example ms1m). For each person in ms1m we compare its center embedding with all 360k and if it does not match with anyone (using a low similarity threshold) we can add that person to the base dataset. You may say that all ms1m identities are included in glint360k, but how about megaface? It has 672k identities.