dash-image-annotator

Image Annotator Web-app using Plotly-Dash, MongoDB, and MLflow

Demo:

Main Task List:

• multi-page web-app
  • image uploader to external MongoDB
    • ingest selected images
    • ingest zipped images
  • annotator
    • query/display images from external MongoDB
    • manual annotation (save to external MongoDB)
      • box
        • editable
      • free-hand lasso
        • opened/closed
      • polygon-lasso (Ctrl+MouseClick)
        • opened/closed
    • DeepLearning model prediction (pyfunc models on MLflow server)
      • one image at a time
      • batch process
• Dockerfile
• Tests
• Jupyter example notebooks:
  • database interaction
  • create train/validation splits with masks
  • upload images with pre-existing masks

Built With:

1. python (version=3.6.7)
2. Plotly-Dash
3. Visdcc
4. MongoDB/GridFS
5. MLflow
6. Pillow

Step-By-Step Guide:

1. $ conda create -n dash_img_app python=3.6.7
2. $ pip install -r requirements.txt
3. set env variables shown in apps/config.py
   1. MLFLOW_URI variable is optional
4. $ pyhton index.py
5. click Navigate to uploader link
   1. drag and drop or select images or zipped images
6. click Navigate to annotator link
7. select number of images
8. select annotation_label_name
   1. optional annotation label-name keyword filter
   2. add ! to keyword for logical NOT
9. click Download button
   1. Select Images datatable will be populated with filenames in MongoDB
10. click a row's radio button in Select Images datatable to display images
11. Select manual annoation buttons for box. free-hand lasso, and polygon-lasso
    1. additional button info:
       1. polygon-lasso requires the Ctrl keyboard button to be held down continuously for each mouse click. release Ctrl button upon completion
       2. only boxes can be edited
          1. toggle Edit boxes True after drawing boxes to make them editable
          2. boxes will appear dotted while editable
          3. toggle Edit boxes False to stop edit mode
       3. toggle Close lasso to connect starting and ending points with a staight line
       4. toggle Open lasso to not connect starting and ending points
12. annotation label names can be set in two ways:
    1. edit Select/Edit Labels datatable labels and colors
       1. select a row's radio button before drawing annotation
    2. click the legend to edit a lable name for a trace and Enter on keyboard
13. traces can be hidden by clicking on their respective legends
14. hidden traces can be removed by clicking the Remove Hidden Traces button
15. annotations can be saved to the MongoDB by clicking the Save button

TODO: add mlflow model instructions

References:

1. Example: Upload and Download Files with Plotly Dash
2. Visdcc