What has changed compared to version 0.5?
In short, everything, from the back-end to the user interface. But ilastik still does the same thing: you train it interactively and it learns to do image processing for you. The different tasks ilastik can now perform are separated into workflows, which in turn consist of applet modules (see developer documentation for more details on the new architecture). With a few exceptions, we ported all the functionality of ilastik 0.5 to the new version, so if you can't find something you are used to, just contact us and we'll be happy to show you how it can be done in ilastik 1.1.
Updates in the workflows of ilastik 0.5
- Pixel Classification:
- The biggest change is that you are no longer limited by the RAM of your machine. Feature computation and pixel label prediction are now done lazily, so that only the field of view is predicted. Computations are cached and the older caches are discarded when the memory limit gets close.
- All objects are now segmented separately, so that they can be be saved, edited and exported individually. This makes for much faster saving time and much smaller project files. We have added uncertainty estimates from Straehle et al. to allow you to place seeds more efficiently. We also simplified the preprocessing interface to make it easier to choose the right settings.
- Object Classification:
- Same as pixel classification, but for objects. You label objects by clicking on them and ilastik tries to classify them using object-level features. The main difference to pixel classification is that this workflow requires a probability map or a segmentation besides the raw data (obviously, you can acquire this map by running pixel classification first). On the other hand, it can use richer features, including shape, object summary statistics and properties of object neighborhoods.
- ilastik 1.1 includes two workflows for tracking: automatic and manual. The automatic tracking workflow implements the Conservation tracking algorithm ([Schiegg et al., ICCV 2013]), which serves to track multiple dividing objects in presumably very large datasets. As in pixel classification, the algorithm expects you to provide sparse annotations to segment the tracked objects and to detect the divisions. In the manual tracking workflow, we provide a convenient interface to track manually. Simple sub-tracks are automatically extracted by ilastik and you only have to link the objects over the ambiguities.
This workflow allows you to count objects in crowded scenes without performing segmentation or object detection first. Under the hood, it implements the algorithm of [Fiaschi et al., ICPR 2012]. As usual in ilastik, you have to sparsely label the background and give the algorithm a few examples of object counts in rectangular bounding boxes.
What has changed compared to versions 1.0.*?
- Numerous bug fixes and stability improvements. Among others:
- Faster project load times thanks to label storage changes
- Better interactive performance thanks to improved feature matrix caching
- Errors captured in a log file that can be sent to the ilastik team for analysis
- Better CPU utilization during batch export
- Many more images can be loaded thanks to threading optimizations
- Conservation tracking instead of chaingraph tracking in automatic tracking workflow
- Classifiers other than Random Forest in pixel classification
- Headless mode also for object classification
- Graph Cut segmentation with Ising model in Thresholding applet
- Mesh export in carving
- Export of results other than probability maps
What has changed compared to version 1.1.5?
- New Workflow: IIBoost Synapse Detection
- Performance Enahancements:
- New Batch Mode Implementation - Faster set-up time for 1000s of images
- Faster Carving Backend
- Improved Memory Management
- New Features:
- Better support for reading multi-dimensional TIFF datasets, including OME-TIFF
- "Camera" export button: Export image views exactly as they appear in the Window.
- Object Classification: New Object Statistics Window
- Object Counting:
- Export to CSV
- Headless-mode support
- Runs on more Linux distros, including CentOS 6
- Numerous bug fixes and stability improvements