Shelf Monitoring: Computer Vision for Retail Shelf Audits

Shelf monitoring automates product placement checks in retail. We build computer vision systems that analyze each shelf every 15 minutes using existing cameras or autonomous robots. According to IHL Group, out-of-stock costs retailers $1 trillion annually globally — about 4% of revenue. Traditional inspections: employee walkthroughs take hours and provide a snapshot. Our automation reduces response time to minutes and increases accuracy to 95%.

Our experience — 12 years in CV for retail and 50+ implementations in Russia and CIS. We use YOLOv8 (accuracy mAP@50 — 0.92) and custom architectures for SKU recognition. In this article, we'll break down how we build such systems: from dataset collection to deployment on cameras or robots.

Why retailers lose billions on empty shelves?

Out-of-stock is not the only problem. Incorrect placement (planogram violation) reduces sales by 3–5%. Products not in their places — customers leave. Price tags mismatch — fines and dissatisfaction. Manual checks are expensive: one inspector on average visits 10 stores per week, spending 20 minutes per shelf. A chain of 200 stores means 400 person-hours per week just for shelf checks. A CV system does the same in 5 minutes per entire store, without interrupting operations. Each out-of-stock for a popular SKU costs on average $25 in lost profit.

How does product detection on the shelf work?

We use YOLOv8 — a state-of-the-art real-time detector. On custom datasets, it gives mAP@50 = 0.92 on typical sets, which is 15% higher than previous versions. The model is trained on images from various angles and lighting conditions. The code below shows a basic class for analyzing shelf images.

from ultralytics import YOLO
import numpy as np
import cv2

class ShelfMonitoringSystem:
    def __init__(self, product_model_path: str,
                 planogram_path: str = None):
        self.detector = YOLO(product_model_path)
        self.planogram = self._load_planogram(planogram_path) if planogram_path else None

    def analyze_shelf_image(self, image: np.ndarray) -> dict:
        """Analysis of one shelf photo"""
        # Detection of all SKUs
        detections = self.detector(image, conf=0.4, iou=0.5)

        detected_products = []
        for box in detections[0].boxes:
            sku = self.detector.model.names[int(box.cls)]
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            detected_products.append({
                'sku': sku,
                'bbox': [x1, y1, x2, y2],
                'confidence': float(box.conf),
                'facings': 1  # each bounding box = 1 facing
            })

        # Aggregation by SKU
        sku_counts = {}
        for product in detected_products:
            sku = product['sku']
            if sku not in sku_counts:
                sku_counts[sku] = {'count': 0, 'positions': []}
            sku_counts[sku]['count'] += 1
            sku_counts[sku]['positions'].append(product['bbox'])

        result = {
            'detected_skus': sku_counts,
            'total_facings': len(detected_products),
        }

        # Comparison with planogram
        if self.planogram:
            result['planogram_compliance'] = self._check_planogram(
                sku_counts, self.planogram
            )
            result['out_of_stock'] = self._find_out_of_stock(
                sku_counts, self.planogram
            )

        return result

    def _find_out_of_stock(self, current: dict, planogram: dict) -> list:
        """Finding missing products"""
        missing = []
        for sku, expected_facings in planogram.items():
            current_facings = current.get(sku, {}).get('count', 0)
            if current_facings == 0:
                missing.append({'sku': sku, 'status': 'out_of_stock',
                                 'expected_facings': expected_facings})
            elif current_facings < expected_facings * 0.5:
                missing.append({'sku': sku, 'status': 'low_stock',
                                 'current': current_facings,
                                 'expected': expected_facings})
        return missing

Training the model on a catalog of 5000+ SKUs

For a large retailer, we use hierarchical classification: first determine the category, then the specific SKU. Each SKU requires at least 500 labeled images. Augmentations account for shelf specifics — rotations, scale, glare.

# Dataset structure: cropped product images by folder
# dataset/
#   train/
#     category_beverages/
#       sku_cola_1l/
#       sku_juice_orange/
#     category_dairy/
#       ...

from ultralytics import YOLO

model = YOLO('yolov8l.pt')
model.train(
    data='shelf_products.yaml',
    epochs=200,
    imgsz=640,
    batch=32,
    workers=8,
    optimizer='AdamW',
    lr0=1e-3,
    # Shelf-specific augmentations
    degrees=5.0,        # small rotation
    scale=0.3,          # scale change (different distances to shelf)
    fliplr=0.5,
    hsv_h=0.02,         # slight color change
    mosaic=1.0
)

Mobile app for inspectors

An employee takes a photo of the shelf via a mobile app, the system instantly shows:

• Green frame: SKU in stock, matches planogram
• Yellow frame: low stock
• Red frame: out of stock

class ShelfInspectionAPI:
    def __init__(self, monitor: ShelfMonitoringSystem):
        self.monitor = monitor

    def analyze_photo(self, image_bytes: bytes,
                       store_id: str,
                       shelf_id: str) -> dict:
        image = cv2.imdecode(
            np.frombuffer(image_bytes, np.uint8),
            cv2.IMREAD_COLOR
        )
        result = self.monitor.analyze_shelf_image(image)

        # Add visualization
        annotated = self._annotate_image(image, result)

        return {
            'store_id': store_id,
            'shelf_id': shelf_id,
            'analysis': result,
            'annotated_image_base64': encode_image_b64(annotated)
        }

Integration with robots for automated inspection

Autonomous robots (Simbe Tally, Brain Corp) patrol the store and photograph shelves. The CV system processes photos in real time, sends replenishment tasks to staff.

How to reduce losses from empty shelves?

Automation allows detecting out-of-stock 10 times faster than manual walkthroughs. The system compares the current state with the planogram and generates tasks for employees. A typical scenario: a robot patrols the store in one hour, the system identifies 15–20 violations, staff receives notifications on mobile devices. Response time — less than 5 minutes. This reduces revenue losses by 80%.

What is included in our work

1. Audit of current processes and cameras — assessment of image quality, lighting, viewing angles.
2. Collection and labeling of dataset — at least 5000 images per category, labeling quality control.
3. Model training with metrics mAP > 0.9 and recall > 0.85.
4. Integration with WMS/ERP (SAP, 1C, Oracle) via REST API.
5. Development of a mobile app for Android/iOS with online analysis.
6. Visualization of results in dashboards (Grafana, Power BI).
7. Documentation and training of the client's team.
8. 6 months of support, including retraining when new SKUs are added.

Timelines and Metrics

For small chains (up to 500 SKUs), implementation takes from 5 weeks; for large ones (5000+ SKUs), up to 16 weeks. We guarantee SKU detection accuracy of at least 88% and compliance check accuracy of 85%.

Our experience includes integrations with WMS SAP and 1C, as well as with Simbe Tally and Brain Corp robots. Certified engineers in PyTorch and TensorFlow. Over 5 years in the market, 12 years of expertise in computer vision.

Want to assess the feasibility of implementation? Contact us — we will conduct a free audit of your cameras and provide a proof-of-concept in 2 weeks. Request a consultation to discuss your project.

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.