Zhen Li
Zhen Li
You could change the `--step` and `--neighbor_stride` for saving GPU memory, but the performance will be degraded. Besides, we are working on a more efficient inference pipeline for E2FGVI-HQ, but...
Could you tell me the spatial resolution of your video？
We use a GPU with 48G memory to process 720p video. The setting of `--step` does affect the performance. To check the influence of this parameter, you could first keep...
很好的问题！ 数据归一化的方式我们follow了STTN和FuseFormer的做法。我们没有尝试过遮挡区域设置为-1的做法，因此无法确定对模型的影响。
您好，数目8是为了公平的与STTN和FuseFormer进行对比。 我们在文章的补充材料中提供了切分后的3D窗口尺寸大小: (T_{l} + T_{nl}) × 5 × 9. 在实验中，我们试过对t进行split，但发现将t设置为与输入序列长度一致的情况效果最好。
Thank you in advance! We will check this PR ASAP.
Sorry for the late reply. You could remove ModulatedDeformConv2dFunction from [model/modules/feat_prop.py](https://github.com/MCG-NKU/E2FGVI/blob/924b56c133fffe37327f9c9b90290fc3d0538581/model/modules/feat_prop.py#L127) and re-train a new model without deformable alignment. But there is a little cost in inpainting performance.
谢谢你的关注。 你的理解没问题，谢谢你指出来。这确实是一个BUG，我们实验了一下，该BUG修复后，PSNR还会上涨约0.1dB（前250k）。
You could download the source code from arxiv.
1. We indeed load all frames and masks into VRAM at once. Your solution is the most straightforward way. 2. We discussed about this case when we wrote the test...