Something wrong with results of video inpainting demo.

[JOyer193]

I used the following command to test.
python tools/video_inpaint.py --frame_dir ./demo/frames --MASK_ROOT ./demo/masks --img_size 512 832 --LiteFlowNet --DFC --ResNet101 --Propagation

Results turn out to be really bad. Even running deepfill1 independently provides better results than video inpainting.

[lukas-blecher]

Hi,

• How many frames are in the frames and masks directories? The more the better is the result.
• What PyTorch version are you using?
• How many propagation iterations does it need to until the process stops?

It has to be one of these because the comand
python tools/video_inpaint.py --frame_dir ./demo/frames --MASK_ROOT ./demo/masks --img_size 512 832 --LiteFlowNet --DFC --ResNet101 --Propagation
still works fine for me.

[JOyer193]

Hi,

• How many frames are in the frames and masks directories? The more the better is the result.
• What PyTorch version are you using?
• How many propagation iterations does it need to until the process stops?

It has to be one of these because the comand
python tools/video_inpaint.py --frame_dir ./demo/frames --MASK_ROOT ./demo/masks --img_size 512 832 --LiteFlowNet --DFC --ResNet101 --Propagation
still works fine for me.

Thanks for your fast reply.

• I used all 80 frames and mask from those directories.
• Pytorch 1.3 and cupy-cuda100
• It seems only 80 frames have been processed within 2 iterations.

[lukas-blecher]

I'm guessing you are using the standard models: sintel, resnet101_movie and imagenet_deepfill?
That's very weird. It should produce the same results I get.
When I run it, it also only takes one propagation iteration to stop.
Did you change any other default settings in the code?

[JOyer193]

I'm guessing you are using the standard models: sintel, resnet101_movie and imagenet_deepfill?
That's very weird. It should produce the same results I get.
When I run it, it also only takes one propagation iteration to stop.
Did you change any other default settings in the code?

It seems like my environment causes the problem. I used exactly the same code in a new environment, It works alright.