Urgent -- Full Dataset Not Downloading

Hello,

I have been using roboflow to annotate a dataset (object detection on MacOS and using Safari) and, after generating the first version of it, cannot download the full generated version. The dataset split for this version is 853 training images, 244 validation images, and 121 testing images. When I click to view all images, only 36 load in. If I click ‘Load More…’, the website starts loading, buffers, and then stops trying to load more images in. These 36 are the only ones that get downloaded across all three splits.

I generated a third version as well and was only able to download 207/1389 total images. I have tried every downloading method offered and none download the full dataset version. Various other aspects of this project have been buggy, such as viewing the full dataset (previously only 8 of ~1200 images would be loaded when I went to view them) and previewing annotated images before adding them to the dataset (the previews would show the incorrect annotations, so I would have to click on each image to ensure the annotations were actually correct).

Here is the link to my dataset on the universe:

https://universe.roboflow.com/hex-nuts/stud-identification

Support Key: KprBpe8543boYFLKjmlJAb9gw802

Without this data, I cannot train and my entire project is at a standstill. Having to re-annotate this data on a different platform would be devastating in terms of time lost. I have also emailed this issue to help@roboflow.com and would greatly appreciate this issue being addressed as quickly as possible.

1 Like

Same issue here, almost a day with the process freezed.

The worst is that you can’t export the labeled data to any other annotation tool like CVAT nor train any model.

Hopefully they’ll fix it soon.

Hey @Rafael_Gildin @Jessica_Tingley

Thank you for bringing this issue to our attention. I’m attempting to recreate your issues but am not seeing any reduced numbers of images.

Could you try generating a new version and downloading?

Hi @leo ,

I just generated a new version and tried downloading but still saw only a fraction of my dataset. You can see that in this new version, the number of ‘total’ images doesn’t match the numbers given under training, test, and validation tabs. However, this lesser number shown (211) is the number of images downloaded.


I exported a zip to my computer using format used by YOLOv8. As you can see, when I check in my downloads I only have 147 training images.

Please let me know if I can be of more help.

Hey @Jessica_Tingley, I noticed that a large portion of your dataset (the part that isn’t working) is composed of images that are in HEIC format. This format will not work with Roboflow – could you try converting those images to jpg or png and re-uploading to a new dataset? You can use a library like pillow-heif · PyPI to do the conversion.

Hi @Lake ,

I reuploaded the images to a new project: Stud Identification2 Object Detection Dataset by Hex Nuts

I had converted the images to .jpeg on my Mac before uploading to roboflow because I got an error when trying to upload them as .HEIC, so I’m not sure how they’re showing as this file type. Are you able to pull the annotations from these images?

Thank you

@Jessica_Tingley - did you actually convert the format, or did you just change the file extension name?

Here’s a guide from Adobe on how to make the conversion: https://www.adobe.com/acrobat/hub/how-to-convert-heic-to-jpg-on-a-mac.html

Hi @Jacob_Witt ,

Thanks for getting back to me. The steps outlined in the link you provided are how I’ve been converting my images (using Preview and exporting as jpg). I uploaded the first part of my dataset back in December/January, maybe I was doing it differently then. Do the images in the new dataset I made look right?

I guess my main question now is whether I can get the annotations from the part of my dataset that isn’t working or whether I’ll have to start annotating over again.

Thanks again,
Jessica

Hi @Lake @leo @Jacob_Witt ,

I’m wondering if there are any updates about this issue and whether the annotations on the ‘bad’ part of my dataset are still viable.

Thank you,
Jessica

Hi @Jessica_Tingley,

Unfortunately, there won’t be any way to recover the annotations for the bad part of the dataset.

I’d be happy to cross check and make sure your new dataset has images formatted correctly. It doesn’t look like any images are in Stud Identification2 Object Detection Dataset by Hex Nuts, could you share the project with Roboflow Support so I can help debug? You can toggle Roboflow Support on the invite page.