Hassan

For „object detection“ experts 😎

I have checked albumentations and seems there is an augmentation method called BboxSafeRandomCrop but does not exactly serve my purpose. Does anyone know something similar that I can reuse?

Therefore, I am looking for a technique where I can only load parts of the images in the data loader, subject to two constraints: image size should be respected and there should be an object in the sample.

Tiling would eventually work, but I might have a lot of negative samples (that would lead to empty iterations …)

Resizing is not an acceptable solution in my case because I would loose a lot of features as the objects are very small (weeds from 12 meters high camera).

Question for experts in OD:
Apart from image tiling and resizing, what are other strategies to feed data where images are of a size 8000x6000 pixels into a data loader for an object detection problem?

Beyond pytorch or tensorflow, what are the tools that make up your computer vision tech stack? What specific tools do you use fore data selection, annotations, MLops, deployement

Insane!

If they want to sue everyone for having got public support to create a product/service to be commercialized, then theoretically, they should sue all private companies contributing to an innosuisse project!