Overview
ISO/IEC 15938-18:2025 - Information technology - Multimedia content description interface - Part 18: Conformance and reference software for compression of neural networks - specifies conformance testing procedures and provides reference software and conformance bitstreams for implementations of ISO/IEC 15938-17. The second edition updates the reference material and bitstreams to cover features added in ISO/IEC 15938-17:2024. The document includes both encoder and decoder reference implementations intended to support interoperability, testing and education for standardized neural network compression in multimedia content description.
Key topics and requirements
- Conformance testing framework
- Procedures for testing decoders and bitstreams (see Clauses 4.2–4.6).
- Tests use supplied test data (reference models and bitstreams) and defined procedures for complete and incremental models, self-contained bitstreams, and bitstreams requiring out‑of‑band parameters.
- Reference bitstreams and models
- A curated set of bitstreams and corresponding source neural network models are provided for conformance testing (distributed as reference_models_bitstreams_ed2_v1.1.zip).
- Bitstreams cover features such as sparsification, pruning, topology pruning and incremental coding; tables in the standard map bitstreams to tested features and encoder configurations.
- Reference software
- Includes encoder and decoder implementations, installation and license information, software architecture and modules.
- Describes parameter reduction methods, parameter approximation, reconstruction, and encode/decode workflows.
- Defines core data structures and interfaces (e.g., model_info, approx_data, nnc_core, nnc_core.nnr_model).
- Annex provides an informative Python® implementation example.
- Test procedures
- Clear procedures for testing decoder conformance and for validating produced bitstreams against ISO/IEC 15938-17 decoding requirements.
Practical applications and users
- Implementers of ISO/IEC 15938-17: software vendors building encoders/decoders can use reference software and conformance bitstreams to validate compliance and interoperability.
- Quality assurance and interoperability labs: run conformance suites to certify decoder/bitstream behavior.
- R&D teams and educators: use the reference implementations and Python annex to learn, prototype and demonstrate standardized neural network compression workflows for multimedia description and analysis.
- Tooling and codec integrators: integrate standardized compressed neural network exchange into content pipelines, edge devices and multimedia services.
Related standards
- ISO/IEC 15938-17:2024 - Compression of neural networks for multimedia content description and analysis (normative reference for conformance and decoding rules).
Keywords: ISO/IEC 15938-18, conformance testing, reference software, neural network compression, conformance bitstreams, reference encoder, reference decoder, multimedia content description.