The most important improvement is that multiclass classification is now fully supported.
For typical segmentation projects the changes should be backwards compatible. If you have overridden one of these configuration options however, it is a good idea to review the documentation in project_defaults.ini to check if there are caveats to watch out for:
train.image_augmentations, train.mask_augmentations
train.classes
Improvements
- Add support for multiclass classification: check out the train.classes parameter in project_defaults.ini
- Add option to use a cron schedule to configure when images can be downloaded: check
out the download.cron_schedule parameter in project_defaults.ini
- Add option to simplify vector output on-the-fly during prediction to make output
smaller: check out the predict.simplify_algorythm and predict.simplify_tolerance
parameters in project_defaults.ini
- Make postprocessing steps to be done configurable in config files: check out the
postprocess section in project_defaults.ini
- Vectorize predictions on-the-fly during prediction to evade need for large temporary storage
- Add end-to-end tests using a test project
- Add option to save augmented images during training for troubleshooting: check out
the train.save_augmented_subdir parameter in project_defaults.ini
- Cleanup support for uninteresting models
- Improve logging, progress reporting, type annotations
- Update tensorflow dependency to 2.4
Bugs fixed
- Fix error in algorithm to determine if label files have changed since last training