patricktum / sen12ms-cr-ts Goto Github PK

I've found a small bug in your download script at line 284. The script effectively only executes wget with -O $filename$ and not with -P $dl_extract_to -O $filename as expected, which is because of this strange behavior of wget.

Thus, the file is downloaded into the current directory and not into $dl_extract_to, and tar throws a "file not found" error for $dl_extract_to'/'$filename, because the file is not there.

A fix would be to wget the file using -O $dl_extract_to'/'$filename at line 284. I tested it and it works.

As title states, can't access the training split.
best regards

Thank you for sharing this code, its very useful. I want to download the pre-trained models, but I noticed that the links are "Not Found".
(For example, the resnet model: https://syncandshare.lrz.de/dl/fiFfN2bj6DaFXfGEGAaAvdZE/baseline_resnet.pth)
Could you please share these models again? Thank you again.

Hello, I downloaded the part of Africa in the SEN12MS-CR-TS dataset for experiments. The training data is 'ROIs1970', '21', the validation data is 'ROIs1970', '40', and the test data is 'ROIs1970', '35'. Only 3 time points are kept under each region. The specific structure is shown in Figure 1. According to this, I only modified the part of the code class SEN12MSCRTS(Dataset). I have two questions:
1: My original intention was to select 3 time points for the experiment at one location, so I modified self.time_points = range(3), but according to the printed results, only the data at time 0 of each location was read. The data of the other two time points cannot be read. Don't know where is the problem, can you give me some advice?
2: I don’t know if it’s because my training set is too small, but the final test result failed. As shown in Figure 2. It's not completely black, and you can still see some outlines when you zoom in. Not sure why this is happening, can you give me some advice?

Hello, I read the paper but I struggle to understand something. The dataset supposedly contains pairs of cloudy and cloudless images. And the cloudy images are not created synthetically. Does this mean the paired images are not temporally aligned?

If yes, doesn't this introduce a major flaw into models trained on this, since it might also learn temporal changes?
I would expect the purpose of cloud removal is to estimate what is under the clouds "right now", not at a different time. For that we could just retrieve the most recent cloudless image. Am I missing something?

Hi, thanks for the amazing dataset and paper!

I have a question about using the checkpoints from this website for I am unsure about the correct model configuration using test.py, specifically, the netG configuration does not seem to be resnet3d_9blocks_withoutBottleneck and the num of input channels is 13.

Could you kindly review the GitHub codebase and suggest the appropriate test commands to use with the checkpoints?

Thank you for your time and kindness.

Hi there,

I'm trying to replicate the conversion of the dataset to hdf5, but as per this line I don't seem to be able to find where the referenced script is located.

In the 《SEN12MS-CR-TS: A Remote-Sensing Data Set for Multimodal Multitemporal Cloud Removal》,"I see the binary cloud masks that matches the S2 dataset in Figure 3. Is this part of the data publicly available? Where can I obtain it?"

Hello @PatrickTUM,

I downloaded the SEN12MSCR dataset for cloud removal tasks. I can understand that s1 has 2 channels and s2_cloudy has 13 channels, but I don't understand why s2 has 13 channels. Why not use the rgb channel directly, which can also reduce the amount of calculation.

Hi! Thanks for the amazing dataset! I followed the instructions and downloaded the entire dataset on my server. However, the total size is around 1.2TB, contradicting to the 2TB reported in your paper. Am I missing any data? Thank you!