search

littlebeen / DDPM-Enhancement-for-Cloud-Removal

A cloud removal dataset and a diffusion-based cloud removal method
29 stars 1 forks source link

Pre-train models #8

Open kequanpan opened 1 week ago

kequanpan commented 1 week ago

I have already used your pre-train models on RICE2 but the output is not what i expected. Can you provide some models that are trained in CUHK datasets and divide the RICE datasets into train datasets and test datasets? Thanks

littlebeen commented 1 week ago

Sorry, I don't fully understand your question. Is the pre-trained model could not remove the cloud in RICE2? The separation of the train and test set of RICE2 is random. I could provide their image ids if you want.

kequanpan commented 1 week ago

Happy to receive your reply. I rethought the problems I encountered while studying your paper. 1.I followed the instructions on GITHUB and successfully ran the code based on the RICE2 dataset. But the outputs are a bit different. In mn2, my PSNR is 29.373, SSIM is 0.8825 , LPIPS is 0.0540.(Compared your outputs:33.637,0.8825,0.0476). In mdsa2, my PSNR is 29.872, SSIM is 0.8869, LPIPS is 0.0491.(Compared your outputs:33.598,0.8842,0.0452). I haven't found the cause of this problem yet. 2.In the usage, THE pretrain CR model Part, the models(.pth) are put tinto guided_diffusion/cloudnet/........respectivily. But in THE Test Part, the code need to put the pre-train model into File'Pre_train', i used the models you uploaded on RICE2 with mn and mdsa(The file suffix is ​​pt). if I want to change the dataset to CUHK, do I need to use the pre-trained model on CUHK? Currently, your page has only uploaded the pre-trained model for RICE2. 3.You mentioned pre-trained models in two places in your description, and I can't distinguish the two pre-trained models well. 4.Does each model have its pre-trained parameters? If not, how can I pre-train each model? I am very interested in the direction of your paper, and I am also in the learning stage. I would be very grateful if you could provide some help.

littlebeen commented 1 week ago
  1. The difference between your result and my result may due to the difference of our test set.
  2. I will upload the pre-train model as soon as possible. But I am quite busy recently. Welcome to push me.
  3. Our pre-train model could be divided into two prat. The reference visual prior model (an end-to-end model trained in the other github project link: https://github.com/littlebeen/Cloud-removal-model-collection) and our diffusion model (trained in this project, please follow the train guidance).
kequanpan commented 1 week ago

Thank you for replyinge me. Recently I studied the method you proposed through the code you shared on GITHUB. And I learned a lot from your approach. At the same time, I am exploring this approach further. I still have a few questions that I haven't answered yet, I will sort them out by the end of this week and send them to you. And I will try to train some end-to end models by myself to see if its results are different.