Notes for researching a new classifier

Mtriage is designed so that, as new types of media analysis become relevant, we can deploy them through relevant analysers. The most well-defined way to expand mtriage's capabilities is to train new classifiers to detect qualities in images, videos, audio, and text. This document is a guide outlining questions to guide the background research that needs be done before a new classifier can be implemented and deployed here.

Prebaked

The easiest way to implement a new classifier capability is to take existing functional code, and adapt it so that it can be orchestrated by mtriage.

• is the quality of interest available in any existing public or otherwise available classifiers? For example, there may be a label or set of labels in ImageNet that capture or approximate the quality.
• do any Github repos exist that implement a relevant paper? How functional are they, and what would it take to adapt them to an mtriage analyser?

Training data

• what datasets or other already-labelled training data exists that captures something of the quality?
• are there ways that you could generate synthetic data for the quality? I.e. using our approach for generating synthetic image datasets. If so, can you find assets that would be helpful-- an accurate 3D model, textures, or approximate 3D models?

Existing Literature

• if a classifier to capture the quality of interest does require custom training and development, are there any papers on arxiv or elsewhere that tell how to train for this (or a simliar) quality?

Any questions or comments please write in the #ml-classifiers channel on our Discord server.