Multi-view object segmentation about stcn HOT 2 CLOSED

That would be interesting but we haven't done any experiments on that. Technically it can "kind of" work because we are not explicitly using temporal smoothness. It might help to use some heuristic to sort the images to minimize the change in appearance between frames and feed the frames as a video. Note that the memory frames are an unordered set. Thus, let's say you have a bunch of images of taken at different azimuths (represented by numbers), you can represent them as different types of sequences:
5->6->7->8->4->3->2->1
5->4->3->2->1->6->7->8
1->2->3->4->5->6->7->8
5->6->7->8 AND 5->4->3->2->1 (as two separate sequences)
with mask propagated from the first frame in the sequence. They should all kind of work -- you can try for yourself on your data.

While something like 1->4->8->... would be bad.
Depending on the density of images that you have, you might also want to decrease the mem_every parameter.

from stcn.

Thank you very much! I'll do some tests!

from stcn.