We put a big effort to rewrite some parts of the labeling engine for easier and better work with results, deserialization, an inner structure and so on. That led us to faster initialization (but we have a lot of space for improvements), more predictable regions operations, more stable per-region labeling, non-breaking region removal.
Features
- Relations visualizations
- Nested labeling improvements
- Granularity fixes - now you can enable granularity per label
- New deserialization mechanics for LSF
- LSF faster initialization,
- More predictable regions operations
- More stable per-region labeling
- Non-breaking region removal
Improvements
- Show errors about content loading and CORS problems
- Detailed logs from ML backend on ML page
- Task navigation improvements with hotkeys on the tasks page
- Minimum formatting of completed_at (@TTouka)
- Maintain ordering in source.json and task.json for sequential sampling (@theconsummatedhruv)
- Move external fonts to LS
- Improving documentation clarity and readability
- Add ML predictions on the tasks page
- Table tag support on import and its example
- Previews for audio and texts on the tasks page
- Don't keep premade predictions in completion + Keep premade predictions if we can't get fresh (@hepaajan)
- Support hostname environment variable to docker
- COCO export support for segmentation (@vkhizanov)
Fixes
- Fixed dates in non-Chromium browsers
- Don't stretch out text object with long lines
- Don't save broken regions (polygons, brushes)
- Fixed per-region audio labeling with predictions
- Update s3 signature version to s3v4 to support more AWS regions (@wdvr)
- Fixes for task validation in cloud storages
- Update pip and six packages to avoid fails while LS install
- Fix storage paths with trailing slashes
- Fix Project.next_task to sort task IDs for "sequential" sampling (@k4200)
Contributors
@john-sandall @wdvr @jeannefukumaru @TTouka @theconsummate @kathoum
@k4200 @hepaajan @vkhizanov