New to this dataset. Need further guidance on the paper and repository especially on the dataset part

[DWSuryo]

Background: first I found Sama-COCO in FiftyOne dataset documentation, then I searched Sama-COCO in Google Scholar. There, I found your paper about COCONut dataset. As I see in the paper, the COCONut dataset has object detection features, which I want to use.

I read further on the paper and repository. As far as I read, the COCONut train dataset has S, B, and L variants. S variant is basically COCO2017 dataset with your annotation methods, B variant adds to S variant with annotations for unlabeled COCO2017 images, then L variant adds combines COCO2017 images with Objects365 dataset.

For the validation, there are relabeled COCO2017 validation version and COCONut version. The COCONut version especially uses Objects365 dataset.

That's what I get from the paper. Now for the data itself, I've seen the issues are apparent from Objects365 dataset (#26 ). As I see, this dataset relies on Flickr so I assume eventually some images there may be deleted from the website.

So, with that, I explore and download the whole dataset annotation files via Kaggle. However, on some cases, there are images with 0 kBs, which are mostly from L version (haven't checked for S or B version yet). Then, I use the Huggingface website especially from your profile page. It has instructions to install, especially for S and B variants, so at least I'm guided from those versions. However, the code hasn't supported the L version (yet?). I tried to modify the download_coconut.py to make --split allow for L version, but somehow it failed especially for the dict and json parts.

With these issues, sometimes I am a bit confused because each source from here, Kaggle, and Huggingface. Sorry if this post is kind of jumbled, so I need further guidance. Therefore, I have somewhat broad questions:

1. Where can I find the object detection version for this dataset instead of instance segmentation (referring to paper)?
2. Is the COCONut-L version supported to download via Huggingface?
3. Since there are some images with 0 kBs in Kaggle, where is the complete one, Kaggle or Huggingface?
4. You mentioned some updated annotations files are from drive folder (coconut_val size and composition #24 ), but somehow, I can't find it. which file/folder you are referring to?

I think that's for now. I may question later. Anyway, nice work on the dataset. I'm intrigued to see the further development

[xdeng7]

Background: first I found Sama-COCO in FiftyOne dataset documentation, then I searched Sama-COCO in Google Scholar. There, I found your paper about COCONut dataset. As I see in the paper, the COCONut dataset has object detection features, which I want to use.

I read further on the paper and repository. As far as I read, the COCONut train dataset has S, B, and L variants. S variant is basically COCO2017 dataset with your annotation methods, B variant adds to S variant with annotations for unlabeled COCO2017 images, then L variant adds combines COCO2017 images with Objects365 dataset.

For the validation, there are relabeled COCO2017 validation version and COCONut version. The COCONut version especially uses Objects365 dataset.

That's what I get from the paper. Now for the data itself, I've seen the issues are apparent from Objects365 dataset (#26 ). As I see, this dataset relies on Flickr so I assume eventually some images there may be deleted from the website.

So, with that, I explore and download the whole dataset annotation files via Kaggle. However, on some cases, there are images with 0 kBs, which are mostly from L version (haven't checked for S or B version yet). Then, I use the Huggingface website especially from your profile page. It has instructions to install, especially for S and B variants, so at least I'm guided from those versions. However, the code hasn't supported the L version (yet?). I tried to modify the download_coconut.py to make --split allow for L version, but somehow it failed especially for the dict and json parts.

With these issues, sometimes I am a bit confused because each source from here, Kaggle, and Huggingface. Sorry if this post is kind of jumbled, so I need further guidance. Therefore, I have somewhat broad questions:

1. Where can I find the object detection version for this dataset instead of instance segmentation (referring to paper)?
2. Is the COCONut-L version supported to download via Huggingface?
3. Since there are some images with 0 kBs in Kaggle, where is the complete one, Kaggle or Huggingface?
4. You mentioned some updated annotations files are from drive folder (coconut_val size and composition #24 ), but somehow, I can't find it. which file/folder you are referring to?

I think that's for now. I may question later. Anyway, nice work on the dataset. I'm intrigued to see the further development

for question 4: coconut_val should be ready for you to explore, tutorials are here. A full tutorial for preparing dataset is coming, thanks for your patience.

[DWSuryo]

Okay. Let me know if there are updates regarding my questions

[xdeng7]

Okay. Let me know if there are updates regarding my questions

try the conversion from masks using some python tools, it looks not accurate enough, I should release the original annotated bounding box data in these two days,