Biologist → AI Engineer building computer‑vision systems for wildlife monitoring.

🐦 WatchMyBirds — real‑time bird detection and species classification from webcams & RTSP streams.

🔗 App · 🧪 Training Pipeline · 🤗 Model on HF

• Realtime object detection (YOLOv8/v11) + species classification (EfficientNet)
• End‑to‑end pipeline: data prep → verification → HPO (Optuna/MLflow) → training → ONNX export → Hugging Face upload
• 29 species, ~10k images; automatic pre‑annotation via GroundingDINO (+ optional SAM)
• Results: OD mAP@[.5:.95] = xx % · CLS top‑1 = xx % (val)
• Tracking: MLflow experiments; model registry with “latest” pointers

Python · PyTorch · Ultralytics YOLO · timm/EfficientNet · OpenCV · Optuna · MLflow · Hugging Face · ONNX

WatchMyBirds is an AI-powered real-time object detection system for birdwatching and wildlife monitoring using webcams, IP-cams & RTSP streams. It supports webcams, IP cameras, and RTSP streams, turning any camera into a smart detection system.

Fully automated training pipeline using modern tools like PyTorch, Hugging Face (--> Optuna and MLflow). End-to-end data preparation & training pipelines for object detection and species classification.

🧩 Key Features

• ✅ Modular and configurable (YAML-based)
• 🧠 State-of-the-art transfer learning (EfficientNet)
• 🔍 Automated hyperparameter tuning with Optuna
• 📈 MLflow tracking for all experiments
• ☁️ Optional Hugging Face model export
• 🧪 Easily extendable to other image classification tasks

🧭 Features Overview

🔎 Zero-Shot Preannotation

• Uses Grounding DINO (with optional SAM) to preannotate bird bounding boxes.
• Reads metadata CSV, processes images in batches with ensemble prompts.
• Applies NMS, nested-box filtering, and optional SAM refinement.
• Outputs COCO-format annotations back into the CSV.

🗂️ YOLO Dataset Preparation

• Reads and filters metadata based on your criteria.
• Generates YOLO-format image and label files.
• Adds augmentation, class balancing, and negative samples.
• Produces ready-to-train dataset with config and split files.

🚀 YOLO Training Pipeline

• Manages environment and TensorBoard logging.
• Loads configs and runs training with checkpointing.
• Exports the best weights to ONNX.
• Bundles final training and inference configs automatically.

🦜 Classification Dataset Preparation

• Creates square image crops per species.
• Handles train/val/test splitting and class balancing.
• Augments data and outputs organized dataset folders with run settings.

🎯 EfficientNet Fine-Tuning

• Sets up mixed-precision training with mixup/cutmix.
• Progressive backbone unfreezing and LR scheduling.
• Tracks metrics via TensorBoard.
• Saves and exports the best weights (PyTorch + ONNX) with full run configs.

🦾 Autonomous Systems & Robotics
🔬 AI-driven Biodiversity Research
🤖 Multi-Agent Systems & AI

🔬 PalaeoMapEurope – A Python-based visualization tool mapping human artifacts and vegetation zones in Europe during the Quaternary period.
📍 Uses geospatial mapping, radiocarbon dating, and vegetation analysis to study human-environment interactions.
📂 Explore the Repository

• LinkedIn: linkedin.com/in/arminfabritzek
• Email: armin@starmin.de
• ORCID: https://orcid.org/0000-0002-0199-6802