Image Dataset Scorer

A simple, graphical, local image dataset viewer which allows you to generate human feedback ("RLHF") metadata, written in Python

The script currently implements the following features and can be easily extended/altered to suit your needs:

• Rate images as either Liked or Disliked
• Easy to interpret, user-friendly interface
• Create or Select a ratings file (JSON) to associate with a particular dataset or objective
• Load altenate datasets on the fly with Load Dataset
• Select a target folder and search it and its subfolders recursively for images
• Filter out previously rated images and enqueue yet-to-be-rated images for evaluation
• Run through queued images quickly using Like, Dislike, Skip, and Back buttons
• Automatically display contents of associated imgfilename.txt metadata file if found, when viewing images
• Ratings are stored in real-time, allowing for completion of rating work in convenient chunks
• Images re-rated after traversing backwards in a session have their ratings data updated appropriately
• To produce a dataset of only "Liked" or "Disliked" images, a Copy [Rating] to Target Dir functionality is included
• When copying rated images, automatically copy metadata (.txt) files with matching names
• Filtering modes: all, portrait, landscape, square

Demo

Rating images:

Save rating data:

Convenient keybinds: