search

Dhananjay42 / crackseg9k

[ECCV W 2022] "CrackSeg9k: A Collection and Benchmark for Crack Segmentation Datasets and Frameworks" by Shreyas Kulkarni, Shreyas Singh, Dhananjay Balakrishnan, Siddharth Sharma, Saipraneeth Devunuri, Sai Chowdeswara Rao Korlapati.
49 stars 6 forks source link

For a more complete dataset #2

Closed Theo-Stats closed 1 year ago

Theo-Stats commented 2 years ago

Thank you very much for your work, which is very useful for the crack detection task. According to you provide the address of "https://doi.org/10.7910/DVN/EGIEBY" I got the dataset. However, it has not yet been classified into the three classes mentioned in your paper and has not yet been included the DINO feature. Could you provide the relevant datasets? thank you very much!!!

shreyask3107 commented 2 years ago

Hi @Theo-Stats, Thanks for notifying us. We will make classifications available in the csv and DINO feature generation code with weights, and the outputs will be shared on our repo. We will try to get this done by the end of the weekend.

Theo-Stats commented 1 year ago

Hi @shreyask3107, I am glad to have your reply.

Have you finished updating the dataset? Can you send me the new link?

Also, I have two new questions that I hope you can answer for me.

  1. I found that the DINO algorithm uses the multi-head attention, and each time there are five output images, but the DINO feature part of the input in your model is single channel, how did you convert it?
  2. Do you use pretrained weights of resnet in your model?
shreyask3107 commented 1 year ago

Hi @Theo-Stats ,

@Siddharth02003 is working on it and we will let you know once we are done. Hopefully within a day or two.

  1. We use one of the five outputs (layer 3). We do not use all of the five outputs because they seem to be correlated and one of them shall do the required job more efficiently.

  2. Yes, we do pretrain when resnet backbone is used.

Theo-Stats commented 1 year ago

Hi @shreyask3107 , I would like to know how you use pretrain weight of ResNet on ImageNet, because the input of your model is 4 channels, while the input of the original ResNet is 3 channels. Thanks!

shreyask3107 commented 1 year ago

@Theo-Stats You can refer to our code in DeepLab, we have essentially loaded all layers except the first layer in such case.

Dhananjay42 commented 1 year ago

Hello @Theo-Stats, we have fixed the bugs in the dataset and have also added a folder called 'Heads' which contains the DINO feature maps for our dataset. Instructions on generating the DINO feature maps have also been added to the repository under the folder 'dino'.

Regards and hoping this helps, Dhananjay.