rcdaudt / fully_convolutional_change_detection Goto Github PK

Hi, I am interested in your work, but I dont know how to generate the change map and how to generate the results in your paper.

Hello, this network is awesome. I've been using your code on my custom dataset but now i am trying to run it on a huge dataset and this code only loads all the dataset on the RAM memory. I already have 64 GB of RAM but for my usage i would need much more RAM.

So i am submitting this issue to ask if there's any way to load data for each batch directly on the disk, so i can use this huge dataset.

Can someone tell me if the patch_size has an impact on the accuracy of training?

hi, thank you for publishing your amazing work here!
I was wondering if you also tested the model withoud using a GPU and if any adjustments need to be done within the code if one wants to do so?

thanks in advance for your help!

Hi. I am interested in ur work. Could u produce more details about OSCD pretain and train code?

What pytoch version should I use?

Thanks for your great work on providing both the dataset and the benchmark code for change detection problem set.

I have a question regarding the code. In the notebook, there is only the train_dataset and test_dataset. During training, the test dataset is used as a validation dataset during training. During training, this validation dataset (test_dataset) is also used to tune the appropriate parameters. Wouldn't the model be kind of overfit to this test_dataset also?

Hi Rodrigo,
Your work is so awesome and I am trying to learn everything. But I am new. When I run the code in GPU, the cost time is similar to the time in CPU, both almost 7 hours. And when I tried to adjust the batch_size from 32 to 128, it was even slower. I am not sure why the results shows like that, could you give some advice?
Thank you so much and best,
Yiming

Hello,i am new to image CD problem ,and i've read your paper then tried your model but it didn't show a good accuracy,could you provide some detail about image pretreatment and what loss function you choose to train the model

Hi, first, thanks for presenting such a nice paper!
After reading your paper, I still can't understand the Siamese architecture. As circled out in the figure above, the input2 branch has the maxpooling and upsampling layers between the encoder and decoder, why doesn't the input1 branch have? Have you tested the performance of the architecture whose input1 branch has the maxpooling and upsampling layers?