ANIMA AKITA — Wave 10 WARDOG

Paper: CST Anti-UAV: A Thermal Infrared Benchmark for Tiny UAV Tracking in Complex Scenes ArXiv: https://arxiv.org/abs/2507.23473 Upstream benchmark repo: https://github.com/PCwenyue/CST-Anti-UAV Wave: 10 WARDOG Focus: UAV/drone defense for Shenzhen Robot Fair

Overview

AKITA is a benchmark-first ANIMA module built around the CST Anti-UAV paper. The paper contributes a thermal IR tracking benchmark, not a new trainable model, so the correct initial implementation is:

1. A CST dataset and metric stack that matches the public benchmark.
2. A YOLO26-based AKITA baseline tracker adapted for tiny thermal UAV targets.
3. Planning artifacts that explicitly separate benchmark reproduction from our YOLO26 rebase and ANIMA deployment path.

The paper reports 220 sequences, 240k+ manually annotated frames, 1.44M frame-level attribute annotations, and a best published CST mSA of 35.92% from GlobalTrack.

Verification Summary

• Paper PDF present locally and fully reviewed.
• Public benchmark repo exists, despite the original local scaffold saying Repo: None.
• Public dataset links exist in the upstream README (Google Drive and Baidu).
• Shared volume currently contains visdrone and uavdark135, but no obvious CST Anti-UAV, UAVDT, DroneVehicle, SeaDronesSee, or internal 1.8M UAV mirror.
• Upstream code is evaluation-focused and Windows-path oriented; it does not provide a reusable training stack or a YOLO baseline.

Quick Start

uv pip install -e ".[dev,serve]"

python -m anima_akita backend-info
python -m anima_akita summarize-dataset --root "/path/to/CST Anti-UAV" --subset test
python -m anima_akita track-sequence --sequence-dir "/path/to/CST Anti-UAV/test/SEQ_NAME"
python -m anima_akita track-subset --root "/path/to/CST Anti-UAV" --subset test \
    --report /mnt/artifacts-datai/reports/project_akita/cst_test.md

You can force the backend with ANIMA_BACKEND=mlx, ANIMA_BACKEND=cuda, or ANIMA_BACKEND=cpu.

Serving

# CUDA image on the GPU server
docker compose -f docker/docker-compose.serve.yml up akita-cuda

# MLX image on Apple Silicon
docker compose -f docker/docker-compose.serve.yml up akita-mlx

# Or run the FastAPI server locally
python -m anima_akita.serve
curl http://127.0.0.1:8910/health

ONNX export

python scripts/export_onnx.py \
    --weights /mnt/forge-data/models/yolo/yolo26m-uav.pt \
    --output /mnt/artifacts-datai/exports/project_akita/yolo26m-uav.onnx \
    --imgsz 640 --opset 17 --simplify

Current Scope

This stage implements:

• Config loading from configs/default.toml
• MLX/CUDA/CPU backend resolution
• CST Anti-UAV dataset indexing from the public benchmark layout
• Benchmark metrics for IoU, center error, success, precision, state accuracy, and per-attribute slicing (OC, OV, SV, TC, DBC, CDB)
• A YOLO26 detector + EMA-smoothed single-object tracker for per-sequence and per-subset inference
• A markdown benchmark report renderer wired into the CLI
• A FastAPI service (/health, /info, /predict/frame, /predict/sequence)
• Docker images for both CUDA (GPU server) and MLX (Apple Silicon) backends
• ONNX export helper (scripts/export_onnx.py)
• ROS2 tracker node wrapper (anima_akita.ros2_node)
• anima_module.yaml manifest for the ANIMA registry

This stage does not yet implement:

• CST dataset acquisition or full training run (gated on the Wave-10 UAV mega dataset build, which is in progress on GPUs 2–7)
• TensorRT FP16/FP32 export (scheduled after the first training run)
• HuggingFace weight upload (scheduled alongside training)

Those are planned in PRD.md, prds/README.md, and tasks/INDEX.md.

Project Structure

project_akita/
├── src/anima_akita/        # AKITA package
├── tests/                  # Unit tests
├── configs/                # TOML configs
├── scripts/                # Dataset check/download helpers
├── papers/                 # Paper PDF
├── prds/                   # 7-PRD build suite
├── tasks/                  # Granular execution tasks
├── ASSETS.md               # Dataset and weight manifest
├── PRD.md                  # Top-level implementation brief
└── NEXT_STEPS.md           # Execution ledger

YOLO26 Rebase

All detector-backed work in AKITA is aligned to YOLO26 through the current ultralytics package and the yolo26*.pt weight family:

• Base model: yolo26m.pt
• Edge variants: yolo26n.pt, yolo26s.pt
• Intended fine-tuned internal weight: yolo26m-uav.pt

The paper itself does not provide a YOLO detector. That baseline is our AKITA adaptation.

License

Research and internal evaluation use only. Respect the original paper, benchmark, and dataset licenses before redistribution.