ANIMA DALMATIAN — Wave 10 WARDOG

Paper: Visual Prototype Conditioned Focal Region Generation for UAV-Based Object Detection ArXiv: https://arxiv.org/abs/2604.02966 GitHub: https://github.com/Sirius-Li/UAVGen Defense Score: 38/50 | Tier: T2 Wave: 10 — WARDOG (War Dog Breeds) Focus: UAV/Drone Defense for Shenzhen Robot Fair

Overview

DALMATIAN currently implements the paper-faithful preprocessing and integration scaffold around UAVGen: visual prototype selection, focal-region planning, label refinement, and YOLO26 dataset handoff.

Current verification verdict: conditionally_verified_cto_review. The paper is real and readable, but the public repo does not yet ship runnable code, and the reported training details conflict between the paper body and the supplement.

Quick Start

uv pip install -e ".[dev]"
python -m anima_dalmatian --backend auto status
python -m anima_dalmatian --backend cpu plan-sample
python -m anima_dalmatian export-yolo26 --output artifacts/yolo26-dataset.yaml
python -m anima_dalmatian make-splits --total 1800000 --output artifacts/splits.json
python -m anima_dalmatian write-shard --output artifacts/demo.jsonl --image-root /tmp/images

Serve

docker compose -f docker-compose.serve.yml up --build
curl http://localhost:9018/health

Build Status

• Phase 1 (deterministic scaffold) — DONE
• Phase 2 (FLUX trainer scaffold) — DONE (training blocked until NIGHTHAWK frees GPUs 2-7)
• Phase 3 (dataset loaders + CLI) — DONE
• Phase 4 (eval + export + serve) — DONE
• Phase 5 (hero + ROS2 node + README) — DONE

Project Structure

project_dalmatian/
├── src/anima_dalmatian/     # Bootstrap code: prototype selection, focal regions, label refinement
├── tests/                   # Unit-test skeleton for the deterministic pipeline
├── configs/                 # Paper and adaptation config
├── papers/                  # Cached arXiv PDF + extracted text
├── repositories/            # Reference repo snapshot
├── prds/                    # 7-PRD ANIMA build manual
├── tasks/                   # Granular task graph for implementation
├── ASSETS.md                # Assets, weights, datasets, status
├── VERIFICATION.md          # Reproducibility audit
├── PRD.md                   # Top-level DALMATIAN product requirements
├── NEXT_STEPS.md            # Execution ledger
└── MODULE_TODO.md           # Implementation checklist

Dual Compute

The runtime resolves mlx, cuda, or cpu via src/anima_dalmatian/device.py. The current initial-stage scaffold is pure Python / NumPy for the paper's deterministic pieces and is safe on both Mac Studio and GPU server environments.

Verification Notes

• Paper PDF fetched and read locally from arXiv.
• Public GitHub repo exists, but code is not released yet.
• Shared dataset volume does not currently expose VisDrone, UAVDT, DroneVehicle, or SeaDronesSee at the expected paths.
• We are adapting downstream training to YOLO26, even though the paper evaluates with GFL and RemDet.

License

Research use only. See paper for original license terms.