Wildlife Species Recognition using Self-Supervised Learning

Problem

Wildlife monitoring using camera traps produces massive unlabeled image data. Manual species annotation is slow and expensive. Build an automated wildlife recognition system using self-supervised learning to learn visual representations from unlabeled camera-trap images and videos, followed by fine-tuning for species classification.

Dataset

• Snapshot Serengeti (too large 200G - 600G)
• Caltech Camera Traps (100G) -> prefered
• iNaturalist (200G)
• Smaller Datasets ?

Methodology

• SSL pretraining using:
  • Contrastive learning (SimCLR / BYOL / DINO)
  • Temporal consistency from burst images
  • Fine-tuning for species classification or embedding-based retrieval

Deployment

• NEW: FastAPI + PyTorch microservices architecture
• Two-service design:
  • Ingestion Service (CPU): Motion detection and frame filtering
  • Inference Service (GPU): PyTorch model inference with batching
• Redis queue for decoupled processing
• PostgreSQL/TimescaleDB for results storage
• Kubernetes autoscaling support
• REST API for easy integration

Quick Start

Setup

# Run setup script
./setup.sh

# Start services
cd src
docker-compose up -d

# Check health
curl http://localhost:8000/health  # Ingestion
curl http://localhost:8001/health  # Inference

Process a Video

# Copy video to data directory
cp /path/to/video.mp4 data/videos/

# Submit for processing
curl -X POST http://localhost:8000/process \
  -H "Content-Type: application/json" \
  -d '{"video_path": "video.mp4"}'

# Or use the example script
python3 example.py

Project Structure

├── src/
│   ├── ingestion_service/    # CPU-optimized video processing
│   ├── inference_service/    # GPU-optimized model inference
│   ├── model/                # Model training code
│   ├── shared/               # Shared utilities
│   ├── docker-compose.yaml   # Service orchestration
│   └── README.md            # Detailed documentation
├── k8s/                      # Kubernetes manifests
├── MIGRATION.md             # Migration guide from DeepStream
├── setup.sh                 # Setup script
└── example.py               # Example usage script

Documentation

• src/README.md: Detailed service documentation
• MIGRATION.md: Migration from DeepStream to FastAPI
• src/model/README.md: Model training guide

Features

✅ Motion Detection: Filters 60% of empty frames before GPU processing
✅ Efficient Batching: Process multiple frames simultaneously on GPU
✅ Horizontal Scaling: Scale ingestion and inference independently
✅ REST API: Easy integration with other systems
✅ Real-time Monitoring: Health checks and statistics endpoints
✅ PostgreSQL Storage: Queryable results with TimescaleDB support