We are core contributors to ISO/IEC 23088-2 (MPEG Feature Coding for Machines), the international standard for compressing neural network feature maps in split inference pipelines.

Since the standard’s inception, our proposals have been adopted across 10 FCTM versions — including L-MSFCv2, lightFCTM, PWD, and NN Inner Codec — advancing the test model from v1.0 to the current v9.0. We attend every MPEG plenary meeting, contributing technical proposals, cross-check evaluations, and editor-level contributions to the normative standard text.

We are pioneering hybrid vision coding — a compression framework that delivers both human-viewable reconstruction and machine task performance from a single bitstream, without the overhead of simulcast (compressing twice for two pipelines).

Two interface paradigms are under investigation:

Feature-Input (FCM extension): Machine vision uses decoded features directly; a lightweight human decoder reconstructs a viewable image from the same features.

Image-Input: Shared information is disentangled and recombined to serve both vision and reconstruction paths with high efficiency.

We are coordinating a joint contribution to MPEG WG2 (October 2026) with international partners (KHU, ZJU, SFU) to formally define use cases, requirements, and common test conditions — and to propose an official Exploration Experiment, the first formal step toward a new international standard for hybrid vision coding.