White papers

Introduction

We document the core concepts that shape our computer vision work, from foundational models to next-generation VLM architectures.

This section sets a scientific baseline before applied research: objectives, terminology, methodology, and operational limits.

• VLM

  Vision‑Language Model (VLM): an architecture that aligns images and language for description, reasoning, and response.

• DNAT

  Deep Natural Anonymisation (DNAT): anonymization by synthetic replacement (faces, license plates) instead of blurring. Preserves visual context while protecting identity.

• CV Foundations

  Computer Vision: visual representations, detection, segmentation, and multi-object tracking for demanding industrial scenarios.

• Gaussian

  3D scene representation for real-time rendering and reconstruction (3D Gaussian Splatting).

• JEPA

  Self-supervised architecture that predicts representations in an abstract space rather than reconstructing pixels (Joint Embedding Predictive Architecture, JEPA).

• I-JEPA

  Image variant of JEPA: predicts representations of image parts from other parts (I-JEPA).

• Languages

  Section in preparation: annotation grammars, domain vocabulary, and data schemas to standardize research.

Tech timeline