When I tried to generate new versions based on my dataset, the images got blanked. Images in the dataset are fine, but in versions are not.
Can anyone tell me how to fix the issue, please? I have also tried reduce the amount of images on the same dataset but still don’t work.
That’s an interesting issue. Is the screen correct in stating that you applied no augmentations whatsoever?
yes, no augmentations applied
Do the images from your dataset originate from a specific source? Potentially with an irregular image encoding?
not sure what is a specific source, but it does not from roboflow.
image format is png, and for label format, i have tried xml and yolov5, but both failed the same.
What I was getting at is if you were exporting the images from a certain software that might have a unique image encoding. Could you upload a sample image that you are having issues with?
There seems to be a bug in Roboflow that has trouble with images that are grayscale encoded, which your images are.
I’ve filed a bug report, and in the meantime, you can convert images into RGB color space using this notebook: convert-color-space.ipynb · GitHub
however, it worked on my another dataset which is also in grayscale. here is the example of the worked dateset
What I was reffering to is not necessarily about the actual image color, but the encoding, the method and format of the image data. The two images you sent uses different encodings on the images that are producing errors.
The one you experienced issues with is a PNG with a
Y color space (which controls the brightness, hence grayscale), as opposed to
YUV color space (which has the ability for colors, but just isn’t using them) that your other image has.
thank you for that code, however, converting images in the code also cause a white blank in the majority regions of the images
Could you share the images you had trouble with? I was able to convert the image you sent previously.
Also, if you don’t mind sharing, what source are you getting your images from? Depending on how you are sourcing your images, it might be easier to get you a fix.