Deep learning engines and helpers

This project contains helpers, scripts and classes for three different engines: fastai, icevision and detectron2. fastai is a good choice for semantic segmentation, classification and regression tasks, whereas icevision and detectron2 are good options for object detection and instance segmentation.

fastai

  • drone_detector.engines.fastai.data contains helpers for working with:
    • geotiff files
    • multichannel images
    • semantic segmentation where the target mask is contiunuous
    • Time series of (multichannel) images
  • drone_detector.engines.fastai.losses contains some custom losses to use for semantic segmentation
  • drone_detector.engines.fastai.augmentations contains helpers for using Albumentations with fastai models
  • CLI utilities for inference

detectron2

  • CLI utilities for batch inference
  • Helpers for training
  • Custom TTA modifications that implement also VFlip
  • Data augmentation for rotated bounding boxes