Dual-Focus Multiscale Attention (DFMA) for Real-Time Object Detection in Mixed Reality

This repository contains the official implementation of DFMA — a lightweight dual-path attention module (multi-scale spatial + channel) that plugs into Ultralytics YOLO (v8+ family) to boost accuracy on hard, low-textured, look-alike objects while preserving real-time performance. The method was presented at IEEE ISMAR 2025.

TL;DR: Drop‐in DFMA blocks + ready YAML + scripts to reproduce training.

---

✨ Features

• Plug-and-play DFMA (and optional MSAF (Multi-Scale Attention Fusion)) layers for YOLO backbones/FPN.
• Ultralytics-native YAML (dfma.yaml) for painless integration.
• Python API and optional CLI; simple training code.

---

📅 Roadmap & Availability

• Data generation code (Blender + hybrid backgrounds): coming soon
• Unity (Sentis) MR demo (scene + scripts): coming soon
• Paper (ISMAR 2025): link coming soon

Watch the repository for releases and updates:

• ⭐ Star the repo to bookmark it
• 👀 Watch → “Releases only” to get notified when we publish code & the paper link

---

🔗 Paper Link (placeholder)

Title: Dual-Focus Multiscale Attention for Object Detection in Mixed Reality (ISMAR 2025)
PDF/DOI: coming soon

📣 Changelog

• v0.1.0 — Initial public codebase (DFMA modules, config, registry, tests).
• Next — Add data-generation scripts and Unity demo; publish paper link.