Production Computer Vision That Works Past the Demo

A typical industrial CV deployment runs $50K-$500K depending on scope, model complexity, edge hardware requirements, and regulatory obligations. In manufacturing quality inspection, documented case studies show 6-18 month payback periods. An electronics manufacturer reduced defect escape rates from 2.3% to 0.1%, saving $1.8M annually in warranty claims. An automotive manufacturer deploying predictive maintenance across 200+ CNC machines saved $3.2M per year. The ROI drivers are not just labor displacement (one CV system replaces 3-8 manual inspectors per shift) but quality improvement: 100% inspection versus statistical sampling catches defects that random checks miss, and the downstream savings from fewer warranty claims, recalls, and customer returns typically exceed direct labor savings by 3-5x.

Both, in sequence. Foundation models (SAM 2, Grounding DINO, Florence-2) are transformative for labeling: they cut annotation costs by 40-60% across every task type, with the largest savings on pixel-level segmentation where manual labeling is most expensive. But foundation model inference is 5-10x slower than a purpose-trained detector. You cannot run SAM 2 at line speed on edge hardware. The production pattern that works is using foundation models to generate training labels cheaply, then distilling to a lightweight task-specific model (YOLO, EfficientDet, RT-DETR) optimized for the target device. The foundation model accelerates your annotation pipeline. The distilled model runs in production. Skipping the distillation step is the most common mistake teams make when adopting foundation models for CV.

Cognex and Keyence command roughly 50% of the industrial machine vision market and cover a wide range of standard inspection tasks. Cognex excels at high-accuracy inspection in semiconductor and automotive. Keyence ships integrated bundles that production engineers can configure without writing code. Cognex's 2025 OneVision platform lets non-experts upload defect images and get an auto-trained model deployed to factory cameras. If your inspection task fits standard categories (surface defects, dimensional checks, presence/absence) with stable lighting and fixed product geometry, buying is faster and cheaper. Custom CV is the right choice when the defect taxonomy evolves frequently, when you need to fuse vision with non-visual process data (vibration, thermal, chemical), when regulations require full model traceability, or when environmental variability exceeds what fixed recipes can handle.

Over 70% of organizations report significant performance degradation within six months of deploying a CV model. Drift comes from three sources: data drift (lighting changes, camera aging, material variation), concept drift (new defect types or product variants), and label drift (evolving quality standards). We build monitoring into the deployment from the start. Statistical drift detection using Population Stability Index (PSI > 0.2 triggers alerts) and KS-tests runs continuously on inference outputs. When drift exceeds thresholds, the system flags affected production windows and can trigger automated retraining workflows or queue human review. The monitoring stack uses Evidently AI for metrics, with dashboards tracking prediction distribution shifts and edge-case encounter rates. We also build the retraining pipeline so the correction loop is hours, not weeks.

Traditional CV pipelines (YOLO, EfficientDet, Mask R-CNN) produce structured outputs: bounding boxes, segmentation masks, class labels with confidence scores. They run in single-digit milliseconds, are deterministic, and deploy on edge hardware. Vision-language models (GPT-4o, Gemini, Claude, Qwen2.5-VL) take an image and a text prompt and produce natural language responses. They excel at visual question answering, document understanding, and anomaly description, but run at hundreds of milliseconds per inference with non-deterministic outputs. Open-source VLMs now perform within 5-10% of proprietary models at 64% lower cost when self-hosted. The practical architecture is hybrid: a traditional detector finds objects or anomalies at line speed, then a VLM reasons about context for the cases that need explanation. We build both components and the integration layer between them.

If your software interprets medical images and its output informs clinical decisions, it almost certainly qualifies as Software as a Medical Device and needs FDA authorization. The FDA cleared 295 AI/ML-enabled devices in 2025 alone, bringing the cumulative total to 1,451, with 76% in radiology. 97% went through the 510(k) pathway. The January 2025 draft guidance introduces the Predetermined Change Control Plan, which allows post-market model updates without new submissions if changes stay within pre-approved parameters. This matters for any AI system that learns or adapts. For teams building medical CV, the regulatory pathway should shape architecture decisions from the start: you need documented uncertainty quantification, performance monitoring across demographic subgroups, and version-controlled training pipelines. Retrofitting these for a submission is far more expensive than building them in.

The EU AI Act's prohibited practices took effect February 2, 2025. Untargeted scraping of facial images from the internet or CCTV is an absolute prohibition with penalties up to EUR 35 million or 7% of global turnover. Emotion recognition in workplaces and educational settings is also prohibited except for medical or safety purposes. High-risk CV system rules become enforceable August 2, 2026, covering biometric identification systems, CV in critical infrastructure, employment-related visual analysis, and CV used in law enforcement or migration. High-risk classification requires conformity assessments, CE marking, risk management documentation, data governance records, human oversight mechanisms, and registration in the EU database. Quality inspection CV in manufacturing generally falls outside high-risk unless it is used for worker surveillance or monitoring. We classify systems against the AI Act's risk taxonomy and build the documentation and architectural features that compliance requires.

Dramatically. Manual bounding box annotation costs $0.02-0.09 per object. Manual semantic segmentation costs $5-15 per label. 3D point cloud annotation runs $6-20+ per label. With foundation-model auto-labeling (SAM 2, Grounding DINO, YOLO-World), you generate initial labels from text prompts, then human reviewers correct errors. Review is faster than labeling from scratch. Mature deployments report 40-60% cost reduction versus fully manual workflows. Auto-labeling achieves 70-85% accuracy, sufficient for most detection and segmentation tasks after human correction. The correction step is non-negotiable: 20-30% of labels have quality issues even with manual annotators, and auto-labeling shifts the quality problem from creation to verification. Active learning then focuses human effort on the edge cases where model uncertainty is highest, maximizing the value of every dollar spent on annotation.