- Time spent is not logged or reported
- Contrast/Inversion buttons toggle on but do not toggle off via mouseclick
- User "admin" can revoke their admin status, this should not be possible
- Prompt to create password for user "admin" accepts blank strings, making it impossible to then log in
- Admin page cleanup
- Centre all columns in admin page
- ?? Reword "pending" and "inactive" tags... new columns ??
- ?? Redesign table lines/style ??
- Remove "classification"/"detection" tabs and related columns from "images" tab
- Reword "Segmentation" as "Masks"
- Segmentation view (main annotation view) cleanup
- Reword "score" to "agreement"
- Paintbrush movement creates horizontal/vertical artefacts
- Colourise the helper boxes for better visibility
- Preferences menu
- Standardise wording of options
- Serve images to users based on which have the fewest annotators
- Add easily configurable customisation for the questions at the end of each image ("difficulty" etc.) in project json
- Add configurable default model preferences in project json
- Add helper tips for each field on Preferences tab after ~1 second mouse hover
- Add admin tab for data statistics visualisation over the whole dataset. E.g. class pie-chart, RF confusion matrix, input dimension importance, etc.
- Add "iris export PROJECT" command to save final versions of masks, and output some python-friendly (e.g. pandas) tables to look at dataset statistics
- Add automatic config checker to spot common logic errors. Currently many typos/errors in config do not lead to a good error message.
- CNN features: Deep CNN feature extraction for spatial information gathering at each pixel. Feature map of CNN will be interpolated back to original resolution. User will be decide how many of the features are used with a slider in the Preferences tab, as using too many would create overfitting problems.
- Lasso tool: Isolates area of image for another RF to be used on. Useful when there are two or more distinct areas in the image which the RF cannot handle simultaneously
- Class heirarchies: User annotates lowest subclasses, but they all sit within superclasses. Different RFs are trained for the different parts of the class hierarchy. E.g. a user annotates multiple types of cloud (cumulus/cirrus/etc.) and multiple types of surface (water/land/etc.), then an RF segments "cloud"/"clear", and separate RFs split those into the subcategories.
- Dissimilarity index for suggested annotations: Based loosely on this paper. Input space searched for data that is least similar to already marked areas. This could work per pixel in a single image, or to look for the most useful image to annotate out of the set of available images.