Open txy00001 opened 7 months ago
Hi @txy00001, regarding your questions:
嗨,关于您的问题:
- 去模糊结果在实际场景中受到高度限制,因为这些模型是使用合成数据集训练的。您可以参考论文 https://arxiv.org/abs/2208.13184 进行分析。
- 我将尽快添加这样的推理演示。
- 对于集成到此框架中的模型,我使用开源检查点对其进行了测试,并在同一数据集上获得了相同的评估指标。但是,我没有检查视觉结果。顺便问一下,用于推理的检查点是一样的吗?
The checkpoints are different, the official checkpoints I used, compared the effect of the checkpoints in the frame, the effect of the checkpoints in the frame is poor, and I look forward to using my own video for reasoning as soon as possible
You can try the REDS/NAFNet-width64.yml in the following source code, I use this as a checkpoint for comparison, maybe you can try, of course you can also modify the model based on the video, output the model for the video, so that the framework will be more useful
Hello, I have three questions I would like to consult, 1. After trying some models, I feel that the effect of video deblurring is not very good, are there any improvement measures? 2. Can't reason about your own video directly? I want to enter a video path directly? 3. Compare the deblurring effect of nafnet and the source code in the framework, I feel that the effect of the source code is much better than the effect in the framework, have you done a test comparison?