ANIMA KANGAL

Wave 10 WARDOG | Tier 1 | Defense Score 37/50
Focus: UAV/Drone Defense for Shenzhen Robot Fair
Paper: MMOT: Drone-based Multispectral Multi-Object Tracking

Overview

KANGAL productizes the reusable core of the MMOT benchmark for ANIMA's UAV defense stack:

• Spectral 3D-Stem — 8-band multispectral fusion (paper S4.1)
• Orientation-aware Kalman — 10D state with rotated IoU association (paper S4.2)
• Angle Head — iterative angle refinement for oriented detection (paper S4.3)
• YOLO26 Native — rebased on Ultralytics Jan 2026, NMS-free
• Dual Compute — CUDA + MLX + CPU backends
• Full Pipeline — data prep, training, inference, evaluation, API, Docker, ROS2

Quick Start

uv pip install -e ".[dev]"

python3 -m anima_kangal info
python3 -m anima_kangal check-assets
python3 -m anima_kangal prepare-mmot --root /path/to/MMOT_DATASET
python3 -m anima_kangal mot-to-yolo-obb --mot-dir /path/to/mot --npy-dir /path/to/npy --out-dir /path/to/out

Layout

project_kangal/
├── ASSETS.md
├── PRD.md
├── NEXT_STEPS.md
├── MODULE_TODO.md
├── prds/
├── tasks/
├── configs/
├── scripts/
├── src/anima_kangal/
├── tests/
└── papers/

Key Constraints

• use YOLO26 only for detector rebasing
• support mlx, cuda, and cpu
• do not assume all upstream benchmark code is production-ready
• treat shared datasets as read-only

Reference Links

• Paper: arXiv 2510.12565
• Upstream repo: Annzstbl/MMOT
• Dataset: Hugging Face MMOT

License

Research and internal prototyping only. Respect the upstream dataset and code licenses documented in ASSETS.md.