IceVision

IceVision from airctic is an awesome computer vision framework which offers a curated collection of hundreds of high-quality pre-trained models for: object detection, keypoint detection, and instance segmentation. In Flash, we’ve integrated the IceVision framework to provide: data loading, augmentation, backbones, and heads. We use IceVision components in our: object detection, instance segmentation, and keypoint detection tasks. Take a look at their documentation and star IceVision on GitHub to spread the open source love!

IceData

The IceData library is a community driven dataset hub for IceVision. All of the datasets in IceData can be used out of the box with flash using our .from_folders methods and the parser argument. Take a look at our Keypoint Detection page for an example.

Albumentations with IceVision and Flash

IceVision provides two utilities for using the albumentations library with their models: - the Adapter helper class for adapting an any albumentations transform to work with IceVision records, - the aug_tfms utility function that returns a standard augmentation recipe to get the most out of your model.

In Flash, we use the aug_tfms as default transforms for the: object detection, instance segmentation, and keypoint detection tasks. You can also provide custom transforms from albumentations using the IceVisionTransformAdapter (which relies on the IceVision Adapter underneath). Here’s an example:

import albumentations as A

from flash.core.integrations.icevision.transforms import IceVisionTransformAdapter
from flash.image import ObjectDetectionData

transform = {
    "per_sample_transform": IceVisionTransformAdapter([A.HorizontalFlip(), A.Normalize()]),
}

datamodule = ObjectDetectionData.from_coco(
    ...,
    transform=transform,
)