Assigned
Status Update
No update yet.
Comments
pu...@google.com
nr...@google.com
Hello,
To assist us in conducting thorough investigation, we kindly request your cooperation in providing the following information regarding the reported issue:
- Has this scenario ever worked as expected in the past?
- Do you see this issue constantly or intermittently ?
- If this issue is seen intermittently, then how often do you observe this issue ? Is there any specific scenario or time at which this issue is observed ?
- To help us understand the issue better, please provide detailed steps to reliably reproduce the problem.
- It would be greatly helpful if you could attach screenshots of the output related to this issue.
Your cooperation in providing these details will enable us to dive deeper into the matter and work towards a prompt resolution. We appreciate your assistance and look forward to resolving this issue for you.
Thank you for your understanding and cooperation.
Description
I try to annotate an area which contains the 0 digit only in many files.
The area / text structure looks identical to the human eye.
In part of the files, the digit is extracted. In the second part of the files files, it does not. The video I attached demonstrate that.
The more severe implication of this behaviour, is that I assume it relates to the areas I select in the "wrongly detected" files as blank areas, which prevents me from training a model when these areas are selected.
I emphasize that I can't train even after filling 0 value on its label.
What you expected to happen:
Being able to select areas with 0 value consistently, and use these areas as a part of the data for training.
Steps to reproduce:
1. Upload the files I attached and annotate the zero value, as I did in the attached video.
2. Upload a few copies of each file until there's enough samples to train a model.
3. Initiate a training session. I foresee you'll receive an error.
4. Remove the labels of the zeros falsely detected as blank, and re-initiate a training session. I foresee that you will not get an error.
Other information (workarounds you have tried, documentation consulted, etc):
The only solution I managed to find is to remove the falsely detected labels, which is sub-optimal.
Couldn't find a reference to this topic in the documentation.