About the bounding box refinement of TAO 5.0

I’ve seen bounding box refinement in data augmentation, but only rotation does.
Can I use other data augmentation methods or do bounding box refinement without data augmentation?

Which network are you using?


You can find more in the “Augmentation Config” section of YOLOv4-tiny - NVIDIA Docs.

Sorry, you may have misunderstood me.
I am currently executing getting_started_v5.0.0/notebooks/tao_data_services/kitti.ipynb,
Bounding boxes can be refined based on rotation. I wonder if there are other data augmentation methods such as flipping, or when I don’t use data augmentation, I can also use TAO 5.0 for bounding box refinement?
Or can the bounding box be refined only by rotation?

Yes, you can refer to Offline Data Augmentation - NVIDIA Docs
and 4.0.2 spec file in

wget --content-disposition ‘https://api.ngc.nvidia.com/v2/resources/nvidia/tao/tao-getting-started/versions/4.0.2/files/notebooks/tao_launcher_starter_kit/augment/specs/default_spec.txt

I only see that rotation can refine the bounding box, can’t other data augmentation methods refine the bounding box?

TAO 5.0 can automatically adjust the bounding box size when rotating, but it does not describe whether other data augmentation methods can automatically adjust the bounding box.
Is it possible to roughly label objects and then automatically adjust the bounding box without data augmentation?

Whether refine box or not is set in rotation configuration. See https://github.com/NVIDIA/tao_dataset_suite/blob/35cb4407205cde5fd19e45b288503b0168c28055/nvidia_tao_ds/augment/config/default_config.py#L26 and

But other data augmentation methods can also refine the bounding box according to nvidia_tao_ds/augment/scripts/generate.py.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.