On whether training will enhance results

[gjhhust]

Since the frames are now picked differently than before, I assume the time embedding should be fine-tunable for training? Has the author tried training yet, I'd like to get some help related to starting training fine-tuning!

[zhyjiang]

SAMURAI is a training-free pipeline. We did the modification on the memory selection which does not require the re-training of the original SAM2