fuy34 / dfv Goto Github PK

thank you for your excellent work. But I can't access to your pre-trained checkpoint Our-FV. Would you mind making it public?

in the ddff12 dataloader
the disp_dist is defined as the uniform sampling from max_disp to min_disp
‘’‘
self.n_stack = n_stack
self.disp_dist = torch.linspace(max_disp,min_disp, steps=self.max_n_stack)
’‘’python
the right is the uniform sampling from min_disp to max_disp because of the defocus image in the dataset?

I followed your instruction and used your processed DDFF12 and FoD5000 dataset， then ran the train.py with stack_num 5 batchsize 20 and use_diff 1, but in the evaluation of DDFF12, I get poor results. I did 2 experiments, In the first experiment, I only used DDFF12 to train the network, I got 7.8e-3 in MSE. In the second experiment, I use both DDFF12 and Fod5000 to train the network, but I got 1.66e-2. And my pytorch version is 1.8 , GPU is 2080