ISCC Experiments

A collection of experiments aiding the development of the ISCC

Note: This repository does not contain any production code

We use the iscc_bench package for automated testing of accuracy and performance of the different ISCC components.

MetaID Benchmark

The main purpose of the MetaID component is to serve as a high level grouping component of the full ISCC. It is created from minimal metadata (title, creators) and supposed to identify an "abstract creation" while being helpfull with data deduplication and disambiguation.

Benchmark Approach

For accurate measurment we define four cases:

• True Positive (TP) Two different sets of metadata for the same work result in the same MetaID

• True Negative (TN) Two different sets of metadata for different works result in different MetaIDs

• False Positive (FP) Two different sets of metadata for different works result in the same MetaID

• False Negative (FN) Two different sets of metadata for the same work result in different MetaIDs

The benchmarking is intended to help us maximize true positives and true negatives while minimizing false positives and false negatives.

For automated benchmarking we need reference data from different sources with a common identifier. The wide availability of bibliographic metadata and the common ISBN identifier is a good fit.

Datasets for Metadata

All datasets with at least ISBN, Title, Creators fields qualify for MetaID testing.

Image-ID Benchmark

Algorithms for testing:

• ahash
• dhash
• whash
• blockhash

Datasets for Image-ID

Audio-ID Benchmark

Algorithms for testing

• Chromaprint (AcousticID)
• Dejavu

Datasets for Music-ID

Video-ID Benchmark

Feature Extraction:

• The MPEG-7 Video Signature Tools for Content Identification CCXF6KNnZRcG9-CTKppFGShGYVr-CDbg3f1taa7iV-CRjEWb4JwxW9V

Datasets for Video-ID