hzwer
hzwer
1. 每次发模型前,肉眼观察几十个 hard case or bad case 是不是有提升,不涉及改训练集 2. 是的 3. 训练时是相同大小 patch,只是从不同增广的图中裁剪出来;动态 patch 我想也可以,但是写起来比较繁琐一点 4. SSIM 看起来会让所有指标略好一点,所以有的时候保留;后一项是我观察到光流没有正则化时,在一些动漫场景会出现奇怪 artifacts,所以稍微处理了一下;这两项不要紧 5. 至少对于 rife 这种模型,我观察扩数据量的增益不明显,我的意思是不要太多把精力花在这上边(如果你只有一个人),因为洗数据也很花时间 6. 这个是从别人那抄的参数,可能他们基于某些统计方法平衡的系数,我不觉得会有很大区别,后者就是深层权重稍大了一点
1. 建议对于帧率比较低的(~30fps) 选连续三帧,帧率高的选间隔大一点 2. 只用 Vimeo-Septuplet,它包含 Vimeo-Triplet 3. Vimeo-Septuplet 是主要的,其它可用可不用,我没有做出很明显的差异
@Qjrun Hi, 感觉上是模型的感受野会更大。可能可以用多次图片上 warp 来替代。
Hello, I know that this type of issue occurs during the process of decoding videos into images, such as converting MP4 to PNG. However, I don't know how to resolve...
你好,后者可能是以前为了跑一些demo的遗留代码;1. 可以这样做 但是不建议把分辨率搞这么小再求光流,0.25x 可能就极限了 2. 不能 会很模糊
To install FFmpeg on Windows, follow these steps: 1. **Download FFmpeg**: - Go to the [FFmpeg download page](https://www.gyan.dev/ffmpeg/builds). - Under the "release builds" section, download the latest version of FFmpeg...
Hello, I have some experience: 1. In my previous experiments, I first tried the selection method of (0, x, 6) and found that the effect was not good; random selection...
My detailed ideas can be seen in https://arxiv.org/abs/2310.17294. Warping features is better than only warping images, refer to [Context-aware synthesis for video frame interpolation](https://www.google.com.hk/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&ved=2ahUKEwjSqM-KiZ2DAxUO8WEKHa0hDFwQgAMoAHoECAUQAw&url=https%3A%2F%2Fscholar.google.com%2Fscholar_url%3Furl%3Dhttp%3A%2F%2Fopenaccess.thecvf.com%2Fcontent_cvpr_2018%2Fpapers%2FNiklaus_Context-Aware_Synthesis_for_CVPR_2018_paper.pdf%26hl%3Dzh-CN%26sa%3DX%26ei%3DNl-CZbrpI9eS6rQPgK-FiAE%26scisig%3DAFWwaeZRimTdAZtGdlsS9CK2-8tw%26oi%3Dscholarr&usg=AOvVaw0ryZJCDoUt3bEDkifevXlZ&opi=89978449) It is an encoder learned by...
You may refer to Superslomo
一个主要区别是把 patch 调大到 224 * 448,还有一些 trick 我不知道怎么表述,我想写在一个新 paper 里。。。