search

Jakaria08 / EESRGAN

Small-Object Detection in Remote Sensing (satellite) Images with End-to-End Edge-Enhanced GAN and Object Detector Network
GNU General Public License v3.0
279 stars 70 forks source link

HR/X4 LR/X4 #6

Closed hhakko closed 3 years ago

hhakko commented 4 years ago

For the data set aspect in this paper, I want to consult you a question. I have seen your code. The data input is LR and HR . LR is generated by sampling under the COWC data set. HR is the original image of CoWC. When testing, the input is generated by sampling under the COWC data set. Shouldn't the image we tested be the original CoWC image? If I want to use other data sets, how to set the training input? The test input should be the original test set map of the data set? Isn't the original intention of our algorithm small object dataset? This question bothers me very much. I look forward to your reply,Thank you

Jakaria08 commented 4 years ago

For test input, we need only low-resolution images: 4x downsampled with 64 to 64 tile size for the COWC dataset. Then we get super-resolved images with detection.

For testing a new satellite dataset with a resolution around 0.6 m to 1 m, you should create tiles of 64 to 64 size for our architecture. You can also use larger tiles but might need a large GPU memory. With the new tiles, you also need to create corresponding .txt file for the annotations (similar formatting of COWC dataset).

For training a new dataset, you always need a high-low resolution image pairs with ground truth bounding boxes for detection. This architecture support 64 to 64 tiles for training. for example, If you have low-resolution images of 128 to 128 tiles, then you can create 512 to 512 super-resolve tiles. In that case, you need to change the fully connected layer of the discriminator here: https://github.com/Jakaria08/EESRGAN/blob/7aaae20752e13566dcf119ff1d1d521d0184f6aa/model/model.py#L371 Thank You.

cl886699 commented 3 years ago

this also bother me a lot，in most detection task, we don't have the high resolution images. we can not downsample the original image, and then detect. it is more meaningful to contrast the detection result of original images to the detection result of super resolution images

spacewalk01 commented 3 years ago

@cl886699 you are right, that is the drawback of super-resolution models. Most of them are trained on images down-sampled from high-resolution images which tend to learn downsampling patterns. So, they don't work on actual low-resolution images.