name: The dataset name DVD, GoPro and REDS
mode: construct the dataset for train or val
sampling: the sampling mode for generating pairs of input frames and gt frames.
overlapping: Where sampling consecutive frames, whether overlapped. True or False.
interval: The interval of sampled consecutive frames.
root_gt: The ground truth root path.
data_type: "imgs"
num_frames: The number of frames are sampled.
augmentation: Augmentaions for input frames.
loader: The configs for construct the dataloader.
sampling: The sample mode generating the input frames and gt frames. "n_c", "n_n", "n_r", "n_l" are the general options implemented in SimDeblur. Taking "n_c" for an example, 'n' before '_' stands for n input frames assigned by "num_frames". 'c', 'n', 'r' after '_' means the relative position relationship between input frames and gt frames. Specifically, 'r' is right, 'c' is center, 'l' is left and 'n' is n frames. The following picture depicts it clearly.
augmentation: The image augmentation for training. At now, it includes RandomCrop, RandomHorizontalFlip, RandomVerticalFlip,RandomRotation90, RandomReverse. Each of them has a Randomness controlled by a probability p (default is 0.5).
num_frames: The number of input frames. Many video deblurring frameworks take 5 blurry frames as input, and single image deblurring alorithms take 1 blurry image for input.
Construct your own deep video deblurring dataset
DVD_Train or DVD_Val
├── IMG_0032
│ ├── GT
│ │ ├── 00000.jpg
│ │ ├── 00001.jpg
│ │ ├── 00002.jpg
| | ...
│ ├── input
│ │ ├── 00000.jpg
│ │ ├── 00001.jpg
│ │ ├── 00002.jpg
| | ...