The contradiction of two parameters "num_maskmem" and "max_obj_ptrs_in_encoder"

[JLU-CMX]

num_maskmem=7,
max_obj_ptrs_in_encoder=16,
the two sets in sam2_base.py are different, but I think should be the same logically, cause each of them they should match one identical history frame in the stage of memeory attention.
who can help me?

[heyoeyo]

It does make sense that they should be the same, but they're used in different ways internally and don't (strictly) need to match.

The num_maskmem value controls the size of the learned position encoding (in time) that the model uses on past-frame memory encoding. You can't change this without re-training the model. It does have the effect of limiting the number of 'past frames' that the model can meaningfully encode.

The max_obj_ptrs_in_encoder isn't a model parameter, so it can be changed without re-training the model. It similarly determines how many object pointers are used in memory updates, but it's more like 'application logic' rather than something to do with the model structure.

As for why they are different, it might be due to the data sizes. Object pointers are very small compared to the past-frame memory, so more of them can be included in memory updates without eating up lots of memory or slowing things down.