javiribera / frame-block-matching Goto Github PK

I ran this code line by line to check each results returned. I noticed that frames stored with imsave are extremely noisy and not correlated with the image.

For instance running:
imsave('frames_of_interest/target.png', target_frm)
imsave('frames_of_interest/target.png', anchor_frm)
imsave('frames_of_interest/predicted_anchor.png',predicted_frm) or
imsave('frames_of_interest/error_image.png', error_image)

Returns frames like this:
target frame
anchor frame
predicted anchor
error image

As consequence, I get this quivers which don't seem to make much sense as the change between frames occurs only at the center, yet the quivers move in most directions: quivers and with:
>>> print('PSNR: %s dB' % psnr)
>>> PSNR: 8.295459574843925 dB

Here are the manual steps that I have taken to generate these:
target_frm = np.fromfile('P-thumbnails-1.png',dtype=np.uint8, count=640*360)
anchor_frm = np.fromfile('P-thumbnails-2.png',dtype=np.uint8, count=640*360)
target_frm = np.reshape(target_frm, (360, 640))
anchor_frm = np.reshape(anchor_frm, (360, 640))
ebma = EBMA_searcher(16, 7, 1, 1) predicted_frm, motion_field = \ ebma.run(anchor_frame=anchor_frm, target_frame=target_frm)

I think some debugging in the code is required. As I have taken the steps outlined to generate the vectors but the returned frames are too noisy.

Btw, I used these two images: Img1 and Img2