GPU-accelerated AI service for face recognition, embedding storage, and similarity search, built using InsightFace, Milvus, and gRPC. This service is designed to match faces between facecam and action-shot photos in real-time using cosine similarity.
- 🔍 Detect faces and extract 512-d embeddings using
insightface(with CUDA/GPU support) - 💾 Store embeddings into Milvus vector database
- 🧠 Perform cosine similarity search to find matched faces
- 🖼️ Draw bounding boxes and upload processed images to object storage (MinIO)
- ⚡ Triggered via gRPC with image URL and file ID
photo-svcsends a gRPC request with a photo URL and ID- AI service downloads the image
- Face detection + embedding extraction using InsightFace
- Embedding is stored into Milvus
- Milvus performs similarity search using cosine distance
- Bounding box is drawn and saved
- Result uploaded to MinIO and returned as metadata
# Celery Task Entry Point
def process_photo_task(photo_id: str, file_path: str):
recognizer = get_face_recognizer()
repository = get_vector_repository()
bounding_boxes, embeddings, original_image = recognizer.process_faces(file_path, photo_id)
if embeddings:
for emb_photo_id, face_id, embedding in embeddings:
repository.store_kameramen_embedding(emb_photo_id, face_id, embedding)
matched_user_id, similarity = repository.search_similar_faces(embedding)
if matched_user_id and similarity >= 0.3:
# handle matched user
pass-
Optimize CPU-bound steps (e.g., cv2.imwrite, drawing bounding boxes)
-
Move image read/save operations to GPU if feasible
-
Add database persistence for matched similarity data (MongoDB or other)
-
Dockerize the service for portability
-
Implement CI/CD (e.g., GitHub Actions)
-
Add unit & integration tests
-
REST fallback or HTTP healthcheck endpoint