Does Sahi involve any training step?

[tbonewmy]

Hi there,
if I have a model A trained on a group of image B. say I want to use Sahi to improvement detection of smaller objects in B. When I use Sahi, does it continues to do extra training on those rescaled image path cutouts before it do inference?

[fcakyon]

Thanks for the question. SAHI is only for inference. You can directly use it with your fine-tuned model with no additional training. However we show in our paper that slicing aided fine-tuning increases inference accuracy: https://arxiv.org/abs/2202.06934

You can also check this repo for slicing-aided fine-tuning config files (for mmdetection framework)