panda-lab / single-image-deraining Goto Github PK

hi！Thank you for your work，but i find your dataset that write on read.me is not complete,even some wrong.So can you release the complete dataset? Thanks

Thanks for your good job at first. From your paper, I learned that the synthetic rain streak dataset should has 2400 pairs images for training, and 200 pairs images for testing. But, the synthetic rain streak dataset that download from your link https://pan.baidu.com/s/11t4XIx6f3CEvmOw2XO9fqQ#list/path=%2F has more than 30k pairs images. Could you tell me which parts are used for training and testing ? thanks very much

It seems that there are wrong coordinates of bounding boxes in RID dataset. When I crop bonding box images, it's not related to the classification labels, Please check it, thank you.

Thank you for providing the reference to each method and code.
I would like to ask you if it's possible to have access to the dataset used for the comparisons of the methods (200 rain strike images), with a different system since baidu is not accessible outside of Cina. Thank you in advance.

could you release the code of NIQE SSEQ and BLIINDS-II. Thanks

Sorry to disturb you, I want to ask why ris has no labels?I downloaded it from Baidu Cloud。

The paper mentions that the images have sizes 368x640 and 480x640, but the images in the dataset have 6 variations of dimensions. The problem is that the bounding box dimensions present in xml file is corresponding to 368x640 and 480x640 only.

The solution for this is to resize all images other than those with 480x640 dimensions to 368x640 and then using the annotations as it is. Please either make changes in your dataset or put this in readme.

• Wrong coordinates of image sizes and bounding boxes in RID dataset. For example, see A-D-5 and A-F-78, which represent for two kinds of errors.
• Some labels of RIS dataset do not have paired images. For example, see 125-20160325_091413-300.
• Images in RID dataset need to be flipped for correct matching.

By the way, I wonder the minoverlap threshold you choose to calculate mAP in the paper.
I use YOLOv5 to compute mAP of a subset of RID dataset, and get results shown below. It seems that I need to choose such a low minovelap threshold to get a mAP similar to results in the paper.

Thank you for providing your datasets. But I find that the BaiduYun access of RIS provides only the images without label (ris.zip). And the GoogleDrive's link says I can not access the file. So can you provide the RIS's label? Thank you again!

I can't find the object labels of RIS both in BaiduYun and Google Drive.

Thanks for your good job, but could you release your traing data,thanks

hi thank you for your work, I'm having trouble accessing the data on the Baidu website. I cannot read Chinese, so it is very difficult to access the data. I'm bit unsure what the procedure is to retrieve the real and synthetic data - not the one used for the object detection section. Could the data be placed on a more accessible site? Thank you for your help

我一直在读您的这篇论文，然后有些名字可能不太能理解，我想问您有这篇论文的中文版本吗，如果有可以分享给我看一下吗。本人邮箱[email protected]感激不尽。

In Section 3.3 you said that RID's rain effect is closest to "raindrops" and that of RIS's is closest to "rain and mist". But in section 4.3 you said that you used algorithm trained for rain and mist case for RID and use raindrop case for RIS. Is this a mistake in writing? Or dose this lead to the unsatisfactory result in section 4.3?