DEF-UAVDETR — Anti-Drone Detection Transformer

UAV-DETR: Real-Time Counter-UAS with Wavelet-Enhanced DETR Paper: arXiv:2603.22841 | Yang et al.

Part of the ANIMA Intelligence Compiler Suite by AIFLOW LABS LIMITED.

Overview

Paper-faithful RT-DETR implementation with four key innovations from the UAV-DETR paper:

• WTConv Backbone — Wavelet convolution preserving high-frequency detail for tiny aerial targets
• SWSA-IFI Encoder — Sliding-window self-attention on the deepest semantic feature map
• ECFRFN Neck — Cross-scale recalibration with SBA and RepNCSPELAN4
• Inner-CIoU + NWD Loss — Hybrid regression objective for small-box stability

11.9M parameters | 83K training images | 3 custom CUDA kernels | 5 export formats

Domain

Defense — Counter-UAS / Anti-Drone Detection

Status

• Paper verified + architecture extracted
• PRD-01: Foundation & Config
• PRD-02: Core Model (WTConv + SWSA + ECFRFN + Decoder + Loss)
• PRD-03: Inference Pipeline (CLI + checkpoint I/O + export)
• PRD-05: FastAPI Service + Docker (3-layer serve, CUDA/MLX)
• PRD-06: ROS2 Integration (node + messages + launch)
• PRD-07: Production (telemetry, runtime limits, benchmark, release)
• ANIMA Infrastructure (anima_module.yaml, serve.py, Docker profiles)
• CUDA Kernels (fused wavelet DWT/IDWT, deformable attention)
• Training on 75K Seraphim dataset (GPU, in progress)
• Export: pth + safetensors + ONNX + TRT FP16 + TRT FP32
• Push to HuggingFace

Quick Start

cd project_def_uavdetr
uv venv .venv --python python3.11 && uv sync
uv run pytest tests/ -v
uv run ruff check src/ tests/

Inference

uv run python scripts/run_infer.py --source image.jpg --checkpoint best.pth

API

# Start FastAPI server
uvicorn anima_def_uavdetr.api.app:app --host 0.0.0.0 --port 8080

# Detect
curl -X POST http://localhost:8080/predict -F file=@drone.jpg

Docker

# Full stack (GPU + FastAPI + ROS2)
docker compose -f docker-compose.serve.yml --profile serve up -d

# API only (CPU debug)
docker compose -f docker-compose.serve.yml --profile api up -d

Training

CUDA_VISIBLE_DEVICES=0 nohup uv run python scripts/train_cuda.py \
  --epochs 100 --batch-size 16 --datasets seraphim \
  > train.log 2>&1 & disown

Architecture

Input RGB [B, 3, 640, 640]
  → WTConv Backbone (S2/S3/S4/S5)
  → Project S5 to 256ch
  → SWSA-IFI Encoder
  → ECFRFN Neck (P2..P5)
  → RT-DETR Decoder (300 queries)
  → Detections [x1, y1, x2, y2, score, class]

License

MIT — AIFLOW LABS LIMITED