Label wrapper

User friendly image bootstraping framework.

Label bootstrapping flow

Label wrapper enables label bootstrapping process:

1. Load first data batch
2. Manually label first batch
3. Train first segmentation model
4. Load second data batch
5. Use first trained segmentation model to predict labels
6. Review labels and merge first and second labelled data
7. train the second segmentation model
8. Repeat steps 4.-7. until out of raw data or review of labels is no longer required.

Technical implementation example

1. Load data into dataset
2. Export html
3. Label
4. Export to json
5. Import json and convert json to tfrecords
6. Train on tfrecords
7. Introduce new data
8. Predict with trained model to tf records
9. Import stored tfrecords and convert to html with labels
10. Review stored labels and export to json
11. Join reviewed json and manual json (from step 4)
12. Repeat 5 - 11 for n times
13. Run out of data to label
14. Measure performance

TODO

• Finnish dual data dataset with gtiff (add test)
• mask to shapefile (geocoded)
• shapefile exporter
• shapefile imporoter?
• example inference step with a pretrained segmentation cnn
• (maybe) constructor should take json and load it in postinit
• (maybe) Add via html tests with js (selenium?)

Thanks

Label editor used is VIA 2.0.6.