itsbava/orcan-visiontrace-gpu

Orcan VisionTrace GPU Service

GPU-accelerated face recognition and FAISS indexing service for high-performance reverse image search.

Features

• Batch Face Embedding Extraction: Process multiple images simultaneously using InsightFace on GPU
• GPU-Accelerated FAISS Indexing: Create high-performance vector indexes for similarity search
• Image Enhancement: Automatic quality improvement for poor quality inputs (CCTV, low-light images)
• High-Performance Search: Fast similarity search with adaptive thresholds
• Scalable Architecture: Optimized for production workloads with automatic scaling

API Endpoints

POST /extract_embeddings_batch

Extract face embeddings from multiple images in parallel.

Request:

{
  "images": ["base64_encoded_image1", "base64_encoded_image2"],
  "enhance_quality": true,
  "aggressive_enhancement": false
}

Response:

{
  "embeddings": [[embedding_vector1], [embedding_vector2]],
  "extraction_info": [{"face_count": 1, "confidence": 0.95}, ...],
  "total_processed": 2,
  "successful": 2
}

POST /create_faiss_index

Create optimized FAISS index on GPU for fast similarity search.

Request:

{
  "embeddings": [[embedding1], [embedding2], ...],
  "dataset_size": 10000,
  "dimension": 512
}

POST /search_faiss

Perform similarity search on FAISS index.

GET /health

Health check endpoint.

Hardware Requirements

• GPU: NVIDIA A10G or A100 recommended
• Memory: Minimum 8GB GPU memory
• CUDA: Compatible with CUDA 11.8+

Performance

• Face Extraction: 10-20x faster than CPU (0.05-0.1s per image)
• Index Creation: 5-10x faster than CPU
• Search Latency: <50ms for most queries
• Throughput: 50+ images per batch

Use Cases

• Reverse image search systems
• Identity verification systems
• Photo organization and management
• Security and surveillance applications
• Digital asset management

Model Details

• Face Detection: InsightFace RetinaFace
• Face Recognition: ArcFace embeddings (512-dimensional)
• Enhancement: Multi-strategy image quality improvement
• Indexing: Adaptive FAISS index selection based on dataset size

Limitations

• Requires high-quality face images for best results
• GPU memory limits batch size for very large images
• Cold start latency of ~30 seconds on first request

License

MIT License - See LICENSE file for details.