chensong1995
commented
2 years ago Hello Ginobili-20,
Thanks for your interest in our work! Also, thank you for posting the very first set of discussions!
- On the REDS dataset, each blurry image corresponds to 14 original sharp images.
- During data preprocessing, we use
cv2.resize()for image downsampling. According to the OpenCV documentation, the default interpolation method is INTER_LINEAR. - To the best of my knowledge, there is only one other open-source event-enhanced deblurring project (eSL-Net) available at the time of paper writing. Unfortunately, the open-source version of eSL-Net omits a lot of experimental details, making it difficult to exactly reproduce its data processing steps. As such, we become the first team that establishes a clear procedure to work with REDS in the event setting. This is a challenging task, and we definitely do not expect to be perfect. As for the image size, I believe it is pretty easy to train another version of E-CIR with a different resolution. Should you need any help on this, please feel free to let me know.
- We do not use any data augmentation, such as random croppings and rotations, during the training phase. To be honest, we have not really tried much on this. You are welcome to experiment with different data augmentation methods and let the community know the results.
- In our experiments, we train E-CIR using three Tesla V100-SXM2-32GB GPUs for approximately 100 hours.
I hope this helps! Let me know if you have further concerns.
Best, Chen
HaoDot
Hi @chensong1995 ! It is a really fantastic work for img deblurring or event reconstruction. Besides, you provided a strong pipeline for deblur or interpolation simulation! THX for sharing! However, there are still a few questions confusing me a lot. Hopefully, you can help me to understand. Thx a lot.
the number of sharp imgs to synthesize blurry img: from both paper in arxiv and code for data preparation, I haven't found the exact number of original sharp imgs and the interpolated imgs between the sharps in your setting. I don't whether you could share it with me. The number of averaged imgs determined the difficulty of deblurring at training stage.
question about the interpolation function for downsampling: I have not found he interpolation function for downsampling. I don't know if I have missed something(lol)
the size of downsampled-REDS: The img size in REDS dataset is 1280 720. So after downsampling 4 times, it should be 320 180. However, in your paper and code, the img size is 240 *180. Of course, you want to generate data as close as real APS frame from DAVIS240, however, I don't know if this operation is necessary. Anyway it is a good tip for me!
crop size at training stage: it seems that you didn't crop the img at training stage, I don't know if you worry about the fact that the receptive field of kernal can easily cover a small size of img. Do you have other issues about this?
question about your GPU: According to your paper ,the batch size you set is 96, it's a really big number, besides the representation of ESL is 40-channel-count, which is 40-dimension-tensor and consumes a huge occupation for my card. I really want to know your GPU.
The questions that i proposed above is too many. Besides, chances are that they are not so important. I sincerely hope that you can reply me when you are available. THX a lot!