Synthetic Data Generation for Computer Vision

Labeling 10,000 images with polygon masks takes months of annotator work and significant expense. Yet the real problem is that a client has 400 labeled photos of defects on a conveyor belt, and a YOLOv8 model trained on this volume achieves [email protected] = 0.51. Synthetic data is not a replacement for real data—it is a tool that bridges this gap without hiring an army of labelers. Our experience shows that a properly designed synthetic generation pipeline increases mAP by 30–50% compared to training only on real data. Labeling cost savings reach 80%, which in a typical project translates to savings of $10,000–$50,000 depending on volume.

More about the concept can be read on Wikipedia.

Generation Methods Overview

How to Choose a Synthetic Data Generation Method?

3D rendering yields more predictable labeling quality than GANs but requires 2–3 times more preparation time. Copy-paste augmentation is 10x cheaper than manual labeling for the same volume.

When Synthetic Data Really Helps and When It Doesn't

Three scenarios where synthetic data provides measurable improvement:

• Class imbalance. On a production line, a rare defect occurs once in 5,000 units. Recall for this class is 0.23. Generating 2,000 synthetic instances via copy-paste augmentation on real backgrounds raises recall to 0.71 without changing the architecture.

• Privacy constraints. Medical images, documents, faces—real data either cannot be used outside a secure environment or the volume is clinically insufficient. GAN or diffusion generation with distribution preservation is a workable solution.

• New task without historical data. The warehouse hasn't been built yet, the robot hasn't been purchased, but the model is needed at launch. Rendering 3D scenes in Blender/NVIDIA Omniverse delivers a pre-trained model ahead of time.

Synthetic data does not help if the domain gap is too large—a model trained on renders will fail on real photos without domain adaptation.

What Tools Are Used for Generation?

3D Rendering

NVIDIA Omniverse Replicator is the most mature tool for CV synthetic data. It allows generating images with automatic labeling: bounding boxes, segmentation masks, depth maps, normal maps—all from a single pipeline. Randomization of materials, lighting, and camera position is built-in. According to the official documentation, generating a scene with 10,000 objects takes 2 days.

import omni.replicator.core as rep

with rep.new_layer():
    # Randomize object position
    cube = rep.create.cube(semantics=[("class", "defect")])
    with rep.trigger.on_frame(num_frames=5000):
        with cube:
            rep.modify.pose(
                position=rep.distribution.uniform((-50, 0, -50), (50, 0, 50)),
                rotation=rep.distribution.uniform((0, -180, 0), (0, 180, 0))
            )
        rep.randomizer.lights()

Blender + Python scripting is a more flexible option for custom scenes. Basic knowledge of the Blender API and a batch rendering script suffice.

GAN and Diffusion Models

Stable Diffusion with ControlNet generates photorealistic images from masks or skeletons. For CV tasks, it is relevant: provide a defect mask, and get texture and lighting variations on that mask.

StyleGAN3 works well for faces and medical images with controlled variability. FID (Fréchet Inception Distance) < 10 is achieved on datasets from 10,000 real samples.

Pix2Pix / CycleGAN perform domain transfer: from synthetic to realistic images. This helps close the domain gap without re-labeling.

Copy-paste Augmentation

The cheapest method for detection—cut a real object and paste it onto different backgrounds. Use the albumentations library with CopyPasteAugmentation or a custom script. With proper blending (Gaussian blur on edges, lighting alignment), mAP gain is 5–15% without any new labeling.

Why Validation on Real Data Is Critical?

The main mistake is to mix synthetic data into validation/test sets. Model evaluation should be done only on real data. Otherwise, metrics look good but the model fails in production.

Metrics for evaluating synthetic quality:

• FID — distribution closeness between synthetic and real data (< 30 is acceptable)
• KID (Kernel Inception Distance) — more robust on small samples
• Transfer via model: train on synthetic, test on real—this is the only honest test

Common pitfalls and how to avoid them:

• Using identical backgrounds — the model overfits to floor texture. Solution: randomize background (HDRI maps, random images).
• Ignoring light physics — shadows and reflections don't match reality. Solution: use path tracing or HDR environment.
• Generating only perfect samples — the model doesn't see capture defects (blur, noise). Solution: add realistic noise and blur.

How Much Synthetic Data Is Needed?

In practice, adding 5–10 thousand synthetic images to 500 real ones often increases mAP by 10–15%. For rare classes, up to 20 thousand may be required. Do not mix synthetic data into validation.

Case Study: Glass Defect Detection

Our client is an automotive glass manufacturer. The defectoscopy task: detection of scratches, chips, bubbles, inclusions, cracks, and cloudiness. Initial dataset: 380 real images labeled with bounding boxes. Classes are highly imbalanced — the "inclusion" class has only 23 instances.

Initial model: YOLOv8m with COCO pretrained weights. Result after fine-tuning on the real dataset: [email protected]:0.95 = 0.38, recall for "inclusion" class = 0.19.

Solution:

1. 3D modeling of defects in Blender: an industrial designer created 12 high-detail models (chips, cracks, bubbles) with different materials and textures.
2. Rendered 8,000 images with randomization: 3 lighting types (top, side, backlight), 5 camera angles, random position on the conveyor belt. Used Cycles engine for photorealism.
3. Copy-paste for rare classes: cut real inclusions from 20 original photos, pasted onto 3,000 synthetic backgrounds with Gaussian blur and color correction.
4. Domain adaptation via CycleGAN: translated the style of renders into the style of real line photos (custom dataset of 200 real frames). This reduced FID from 45 to 22.
5. Final dataset: 380 real + 11,000 synthetic. Split 80/10/10 only on real data for val/test.

YOLOv8m hyperparameters:

• batch size = 16, imgsz = 640, optimizer = AdamW (lr=1e-3)
• mosaic augmentation, mixup (0.2), copy-paste (0.5)
• early stopping patience = 10 epochs, max epochs = 100
• trained on single NVIDIA A100 (80 GB) — training took 4.5 hours

Results on real test set:

Savings: the cost of synthetic generation (server time + designer) was several times lower than manual labeling of 11,000 images, reducing the dataset preparation budget by an estimated $20,000.

What's Included

• Initial data audit and problem statement
• Optimal generation method selection (3D rendering / GAN / copy-paste)
• Pipeline development with automatic labeling
• Domain adaptation to reduce domain gap
• Quality validation on real data
• Integration into your MLOps pipeline
• Documentation and team training

Timeline

Cost is calculated individually based on volume and complexity. Our experience: 5+ years in Computer Vision, 20+ projects in synthetic data. We guarantee transparent reporting and post-implementation support.

Get a consultation for your project — we will help select the optimal generation method and estimate timelines. Contact us to assess your task. We will prepare a commercial proposal within one business day.

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.