By
@vyonnie
A media annotation pipeline for images, videos, and audio datasets, with automated quality control using AI. Built for teams consisting of admins, labellers, and QC.
1h 1m 53s logged
Deployed the whole thing onto Render to bypass local installs.
I made a media annotation pipeline for images, videos, and audio datasets, with automated quality control using AI.
There are four stages:
Admin: Create jobs by uploading media URLs and defining category lists.
Labeller: Annotate media tasks with category labels and free-text descriptions.
AI QC: Automated quality control using Gemini to sample and verify labels. Gemini reviews a sample of labels per job:
Human QC: Review flagged labels, approve correct ones, or send them back for relabelling by setting the task status back to open.
Hope this is useful!
1h 45m 14s logged
Going to ship now :) I hope someone can find this helpful!
2h 30m 21s logged
Finished up unit tests for labeller, ai qc and human qc. Also finished the ai qc and human qc endpoints + added human qc html page :)
The image attached here is also a recent addition, the main index.html page showcasing a brief intro of this project, the pipeline, and links to the three html pages.
Also major development: added a way for others to test my project with/without API keys for Gemini or Turso (database), and updated README with implementation steps; will ship soon yayy
2h 36m 46s logged
Added unit tests, restyled both admin and labeller html pages. Started on AI logic
2h 9m 8s logged
Added a text area for labellers to input their names, and updated the “submit labels” button so that it now collects labels from ticked checkboxes and other descriptive labels, and sends them to the database.
So now the labeller’s name and the labels can be recorded in the database! Next step is to integrate AI for checking.
2h 0m 55s logged
Admin page (and connected to backend) done, where “admins” insert urls to be labelled as requested by clients and identify possible categories all urls might fall under, and submit urls and categories to the database. Labeller page partially done, where labellers “get a new task” that renders a url and lists categories as checkboxes to be ticked off and an additional textbox for any additional identifiable labels. To-do next: connect label submission buttons to endpoint in backend.
