Production Face Recognition System Development with ArcFace and FAISS

At a factory entrance, a camera captures an employee's face at an acute angle—the system returns a false rejection, the worker queues up, and security switches to manual verification. 80% of such incidents stem from suboptimal pipeline: the model fails under side lighting or the similarity threshold is set without accounting for real embedding variance. We design production face recognition systems that operate reliably under harsh conditions—from turnstiles to searching among millions of faces on city cameras. Over 5+ years, we have deployed 20+ systems with a False Acceptance Rate (FAR) below 1e-5 and a False Rejection Rate (FRR) under 5% on real customer data. Payback period is less than one year through access automation and reduced security workload.

Full Pipeline

import cv2
import numpy as np
from insightface.app import FaceAnalysis

class FaceRecognitionSystem:
    def __init__(self, db_path: str, threshold: float = 0.5):
        # InsightFace объединяет детекцию + alignment + embedding
        self.app = FaceAnalysis(
            providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
        )
        self.app.prepare(ctx_id=0, det_size=(640, 640))

        self.threshold = threshold
        self.face_db = self._load_database(db_path)

    def identify(self, image: np.ndarray) -> list[dict]:
        faces = self.app.get(image)
        results = []

        for face in faces:
            embedding = face.embedding  # 512-dim ArcFace embedding
            match = self._search_database(embedding)
            results.append({
                'bbox': face.bbox.astype(int).tolist(),
                'person_id': match['id'] if match else None,
                'person_name': match['name'] if match else 'Unknown',
                'similarity': match['similarity'] if match else 0.0,
                'verified': match['similarity'] > self.threshold if match else False
            })
        return results

    def _search_database(self, query_emb: np.ndarray) -> dict | None:
        # Cosine similarity поиск
        similarities = np.dot(self.face_db['embeddings'], query_emb) / (
            np.linalg.norm(self.face_db['embeddings'], axis=1) *
            np.linalg.norm(query_emb)
        )
        best_idx = np.argmax(similarities)
        best_sim = similarities[best_idx]

        if best_sim < self.threshold:
            return None
        return {
            'id': self.face_db['ids'][best_idx],
            'name': self.face_db['names'][best_idx],
            'similarity': float(best_sim)
        }

Choosing an Embedding Model

ArcFace (InsightFace) is the industry standard. LFW accuracy: 99.83%, IJB-C TAR@FAR=1e-4: 96.5% (see InsightFace Benchmark). Embedding size: 512 dimensions. FaceNet (Google) is an earlier model still popular. LFW: 99.65%. Embedding size: 128 or 512 dimensions. MagFace is an improved ArcFace with a scalable margin. IJB-C: 97.1%. For edge devices: MobileFaceNet—1MB, runs on mobile, LFW: 99.5%. Model selection depends on the task: for maximum accuracy choose MagFace, for speed on mobile—MobileFaceNet.

Scaling the Face Database

For small databases (< 10k faces), brute-force cosine similarity works instantly. For large databases—approximate nearest neighbor (ANN). FAISS IVFFlat: search among 1M faces in < 1ms on CPU—1,000 times faster than brute force for databases from 100k faces. The architecture allows horizontal scaling: adding new nodes without service interruption.

import faiss

class FaceDatabase:
    def __init__(self, dimension: int = 512):
        # FAISS IVF индекс для million-scale баз
        quantizer = faiss.IndexFlatIP(dimension)  # Inner Product = cosine sim
        self.index = faiss.IndexIVFFlat(quantizer, dimension, 100)
        self.index.nprobe = 10  # качество vs скорость поиска

    def add_faces(self, embeddings: np.ndarray):
        # Нормализуем для cosine similarity через IP
        faiss.normalize_L2(embeddings)
        if not self.index.is_trained:
            self.index.train(embeddings)
        self.index.add(embeddings)

    def search(self, query: np.ndarray, k: int = 5):
        faiss.normalize_L2(query.reshape(1, -1))
        similarities, indices = self.index.search(query.reshape(1, -1), k)
        return similarities[0], indices[0]

Handling Low-Quality Images

A real-world system must handle blurry, partially occluded, and poorly lit faces. Before identification, we assess face quality using FaceQNet or BRISQUE. Images below the threshold are rejected. Additionally, we employ anti-spoofing (MiniFASNet, CDCN) to protect against photographs and screens. For maximum security, we use 3D liveness detection via IR camera or depth sensor.

Anti-Spoofing: Protection Against Fakes

Face Anti-Spoofing (FAS) is mandatory in production systems. Without it, an attacker could present a photo or video. We integrate MiniFASNet—a lightweight model that runs in real time. For high-security scenarios, we add 3D verification. Savings from preventing incidents can reach 2 million rubles per year for a large facility.

Legal and Ethical Considerations

A face recognition system must comply with legislation: GDPR in the EU, 152-FZ in Russia. Biometric data is a special category of personal data. Mandatory: explicit informed consent, encryption of the embedding database with AES-256-GCM, access logging, and the right to deletion. We include legal auditing in every project.

Additional security measures: role-based access control, regular log audits, ISO 27001 certification.

What Is Included in the Work?

• Requirements analysis: 1:1 verification or 1:N identification, database scale, target hardware.
• Collection of a test dataset from real conditions (lighting, camera angles, cameras).
• Selection and tuning of the embedding model and anti-spoofing.
• Building the database and tuning the similarity threshold.
• Integration, load testing, FAR/FRR monitoring.
• Documentation and operator training.
• Post-launch support.

Development Stages

1. Requirements audit
2. Model selection and dataset collection
3. Pipeline prototyping
4. Integration with hardware
5. Testing (FAR/FRR, latency p99)
6. Deployment and monitoring

Why Choose Us?

5+ years of Computer Vision experience, 20+ deployed face recognition systems. We guarantee: FAR < 1e-5 with FRR < 5% on the customer's test dataset. Contact us for a consultation on your task. Order turnkey development—we will evaluate your project within 2 days. Our technology stack: InsightFace and FAISS.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.