Running inferences backward in eval.py?

[jama1017]

Hi!

In the interactive demo, we can choose to propagate a mask backwards, which can be very helpful for some shots where an object is first zoomed in and then slowly zoomed out. Can we do this in eval.py? (I guess a manual way to do this is to process the video into frames backward, but I just wonder if there is an integrated way)

[hkchengrex]

It is not implemented.

[hkchengrex]

@jama1017 Might be easy to do by just reversing the sorted lists in

XMem/inference/data/video_reader.py

Lines 38 to 40 in af90122

	self.frames = sorted(os.listdir(self.image_dir))
	self.palette = Image.open(path.join(mask_dir, sorted(os.listdir(mask_dir))[0])).getpalette()
	self.first_gt_path = path.join(self.mask_dir, sorted(os.listdir(self.mask_dir))[0])

[jama1017]

Ah that's a good tip. Thanks so much!