Search before asking
Description
Hi!
We are looking to implement RF-DETR as part of a model training pipeline for defect detection. When exploring the code and documentation we have found that a standard resize and crop is applied on the input data. This cannot be disabled by setting aug_config={}.
The issue we are facing here is that we classify whole objects based on their defect. This cropping step risks removing the defect from the image, thus introducing noisy data samples. We already have a custom augmentation step that is equipped to handle this issue.
Would it be possible to provide a simple configuration parameter that allows us to disable this default behaviour?
Thanks in advance for considering this!
Use case
No response
Additional
No response
Are you willing to submit a PR?
Search before asking
Description
Hi!
We are looking to implement RF-DETR as part of a model training pipeline for defect detection. When exploring the code and documentation we have found that a standard resize and crop is applied on the input data. This cannot be disabled by setting
aug_config={}.The issue we are facing here is that we classify whole objects based on their defect. This cropping step risks removing the defect from the image, thus introducing noisy data samples. We already have a custom augmentation step that is equipped to handle this issue.
Would it be possible to provide a simple configuration parameter that allows us to disable this default behaviour?
Thanks in advance for considering this!
Use case
No response
Additional
No response
Are you willing to submit a PR?