Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

请问如果用于多个视频的拼接,需要怎么改造? #291

Closed
qijiaojiao opened this issue Nov 23, 2022 · 4 comments
Closed

请问如果用于多个视频的拼接,需要怎么改造? #291

qijiaojiao opened this issue Nov 23, 2022 · 4 comments

Comments

@qijiaojiao
Copy link

论文的扩展里有提到这个应用,请问代码上怎么改造呢

@hzwer
Copy link
Owner

hzwer commented Nov 23, 2022

在 model/IFNet_m.py 里面把 timestep 改写一下

timestep = (x[:, :1].clone() * 0 + 1) 
n, c, h, w = x.shape
_, XXs = np.mgrid[:h, :w]
XXs = torch.from_numpy(XXs[:,:,np.newaxis]).permute(2, 0, 1).float()[None].to(device)
timestep = XXs * timestep / XXs.max()

这样应该就创建了一个渐变的 temporal encoding

@qijiaojiao
Copy link
Author

在 model/IFNet_m.py 里面把 timestep 改写一下

timestep = (x[:, :1].clone() * 0 + 1) 
n, c, h, w = x.shape
_, XXs = np.mgrid[:h, :w]
XXs = torch.from_numpy(XXs[:,:,np.newaxis]).permute(2, 0, 1).float()[None].to(device)
timestep = XXs * timestep / XXs.max()

这样应该就创建了一个渐变的 temporal encoding

您好,我用rife_m,修改了ifnet_m中如上您给的代码,然后参照论文Fig8做的图片I0 I1(基于demo里的两张图片进行了左右侧裁剪),出来的中间帧图片是一片混乱的像素,可以帮忙看下什么问题吗?
rife_m推理如下:
img_list = [img0]
p = 2 ** args.exp
for i in range(p-1):
t = (i+1)*(1./p)
print(t)
img_list.append(model.inference(img0, img1, timestep=t))
img_list.append(img1)

@hzwer
Copy link
Owner

hzwer commented Nov 28, 2022

我试了一下,用目前的 RIFEm 模型,对这两张图片做插帧会出一片蓝色,具体原因可能是模型无法适应镜头移动,我猜测训练的时候如果加一些镜头移动的增广能修复,很抱歉我之前没有注意到

改用 practical RIFE 里的模型,用 v4.6 模型 https://github.com/hzwer/Practical-RIFE#usage ,可以得到正常的结果

输入(左右各裁剪 32 pixel):
img0
img2

输出:
img1

我的 timestep 生成方式:

n, c, h, w = x.shape
XXs, YYs = np.mgrid[:h, :w]
XXs = torch.from_numpy(XXs[np.newaxis, np.newaxis, :,:]).float().to(device)
YYs = torch.from_numpy(YYs[np.newaxis, np.newaxis, :,:]).float().to(device)
timestep = YYs / YYs.max()

@qijiaojiao
Copy link
Author

谢谢了!

@hzwer hzwer closed this as completed Dec 1, 2022
# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants