From an SE perspective, make a flatmap function that explicitly handles Collection<Optional<Walrus>>. The implementation doesn't matter. If your language/framework already has a compatible flatmap function, make a single frobnicate(Optional<Walrus>) function that returns whatever value is necessary for flatmap(frobnicate) to discard them.
From a CS perspective, doing a filter from Collection<Optional<Walrus>> to Collection<Walrus> is probably a bad idea. If your collection is small, nothing matters. If your collection is large, you probably don't want to spend time making a new copy of it. If your filter just returns a view rather than a hard copy, then there is no optimization benefit and you should just do whatever makes the most sense from an SE perspective. If frobnicate is cheap then you're paying the branch prediction failure tax anyway regardless of when you frobnicate, and if frobnicate is more expensive then your should probably parallelize and have each thread handle unpacking the Optional. Either way, you probably don't want to spend time making a copy.
These are all generalizations based on hypotheticals and there are certainly a lot of exceptions, but broadly speaking I don't see a strong argument here. If optimization matters then optimize based on your own profiling of your situation, and if optimization doesn't matter then design your functions based on what features and paradigms are available/common in your area.