Building a Face Detection System for Production

Building a Face Detection System for Production

Face detection is the first and critical step in any face pipeline. The task is to find all faces in an image and return bounding boxes with confidence scores. At first glance, it seems simple, but real-world conditions—small faces at a distance, profile angles, partial occlusions, poor lighting, masks—turn it into a non-trivial engineering challenge. In production systems, failing to detect even one face can critically degrade the entire pipeline's quality. For example, in stadium video surveillance, faces 20×20 pixels in size are the norm. Without specialized optimization, such objects are missed in 40% of frames. We develop detectors that work in production with latency as low as 4 ms on GPU while maintaining high accuracy even on complex scenes.

Why Standard Detectors Often Fail

Most open-source detectors are trained on datasets like WiderFace, where faces are well-lit and large. In reality, surveillance cameras, outdoor conditions, masks, and glasses reduce accuracy to 60–70%. We solve this by fine-tuning on target data with augmentations that simulate real conditions—rotations, shadows, blur. For example, adding synthetic masks during fine-tuning improves AP from 65% to 89% on the MAFA dataset.

How We Solve Face Detection

We use three main approaches depending on requirements.

SCRFD (Sample and Computation Redistribution for Face Detection) — currently the best speed/quality trade-off. SCRFD-10GF achieves 95.2% AP on WiderFace Hard, which is 2x faster than RetinaFace-R50 with comparable accuracy. More details can be found in the InsightFace repository.

RetinaFace — a classic with landmark detection (5 points: eyes, nose, mouth corners). Used for alignment before face recognition.

YOLOv8 fine-tuned on WiderFace — a versatile option for custom requirements.

from insightface.app import FaceAnalysis
import cv2

# InsightFace: detection + landmark detection
app = FaceAnalysis(allowed_modules=['detection'])
app.prepare(ctx_id=0, det_size=(640, 640))

def detect_faces(image_path: str) -> list[dict]:
    img = cv2.imread(image_path)
    faces = app.get(img)

    results = []
    for face in faces:
        results.append({
            'bbox': face.bbox.astype(int).tolist(),     # [x1, y1, x2, y2]
            'confidence': float(face.det_score),
            'landmarks': face.kps.astype(int).tolist()  # 5 keypoints
        })
    return results

Small Face Detection

Standard detectors miss faces smaller than 16×16 pixels. For surveillance cameras with large distances:

• Image tiling: split the image into overlapping tiles, detect on each, merge results via NMS
• SAHI (Slicing Aided Hyper Inference) — automatic tiling with merge. Library available on GitHub.

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

model = AutoDetectionModel.from_pretrained(
    model_type='yolov8',
    model_path='face_detector.pt',
    confidence_threshold=0.3
)

result = get_sliced_prediction(
    image='crowd.jpg',
    detection_model=model,
    slice_height=512,
    slice_width=512,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2
)

Performance on Different Hardware

How to Fine-Tune a Model for Masked Face Detection?

The pandemic created a separate class of tasks—detecting faces with medical masks. The MAFA dataset contains 35,806 annotated masked faces. Fine-tuning a standard detector on MAFA+WiderFace: AP on masked faces improves from 65% to 89%. The fine-tuning process includes:

1. Collecting or generating synthetic data with masks
2. Augmentation: rotations, lighting changes, blur
3. Fine-tuning a pre-trained model on the mixed dataset
4. Validation on a separate test set

This ensures stable operation with masks, glasses, and other occlusions.

What's Included in Our Face Detection Service

We provide a turnkey solution, including:

• Analysis of your conditions and preparation of synthetic/real data
• Selection and fine-tuning of the detector (SCRFD/RetinaFace/YOLOv8)
• Latency and memory footprint optimization (INT8 quantization, ONNX Runtime)
• Integration into your pipeline (REST API, gRPC, RTSP)
• Documentation and training for your team
• Support during operation

With 5 years of experience in computer vision, we have completed over 30 face detection and recognition projects. We process up to 100 FPS on a single GPU. Results are guaranteed—if accuracy does not meet targets, we refine at no extra cost.

Development Timelines

Request a demo version of the detector for your data and get a preliminary estimate within 1 day. Contact us to discuss your case.

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.