Local Datasets

• Make it more configurable ("bring your own data") with minimal examples / default data to get a grasp of the pipeline

Generate images:

• Generate images to complete our dataset
• Consolidate saved metadata
• Release data with permissive license
• Extension possible : explorer l'influence de la richesse sémantique

Run inferences

• run pipeline: image -> transformation -> compute detection + CLIP features + quality metrics on dataset and add to db
  • compute on original images and images with transformations:
    • use all basic already implemented transformations (resize, filter, text, ...)
    • add one or two combinations aiming at reproducing social media content (TV content if relevant)
• switch to albumentations for transformations
• Consolidate database / ensure that we can properly merge database
• benchmark / evaluation pipeline : break down current huge try-except clause (keep track of failing images / inferences)

Analysis

• Explore results with notebooks
• Consolidate main metrics (eg AUC if relevant) and graphs (eg box plot of perf on different subparts of datasets) as scripts (after handling of db)

Visualisation

• visualisation app (show pre-calculated inferences results + live inference)

Reproducibility :

• add Dockerfile

Doc :

• Include diagrams to explain doc architecture
• Contributing : add MR template for adding new detector
• Contributing : consider adding broader guidelines and explaining current status of project
• Consider using Sphinx doc (and updating current repo accordingly) with Gitlab Pages