AI System for Stores: Item Recognition without Checkout

Item-in-Hand Recognition: A System for Checkout-Free Stores

We've encountered requests for a CV-based checkout-free store that detects item picks and tracks shoppers. Our computer vision in retail system leverages item-in-hand recognition and SKU-level recognition for checkout-free stores. It relies on hand-object detection and metric learning in retail. This builds on Amazon Go technology. AI cameras in store ensure accurate tracking.

Amazon Go solved this problem over 4 years before commercial launch — we offer a realistic pipeline for medium-sized retail. With over 5 years of experience in retail CV and 10+ pilot projects, we have been serving clients since 2018. Accuracy guarantees are backed by pilot projects. In this article, we'll break down how to build such a system from scratch. We cover hand-object contact detection to SKU-level identification in an assortment of thousands of items. We use modern neural network architectures, MLOps tools for production operation, and metric learning for efficient assortment scaling. The result: a 30–50% reduction in losses from theft and inventory errors. Typical savings for a medium-sized store: 1.5–3 million rubles annually, with system pilot costs starting from 500,000 rubles. Our system pays for itself in 6–12 months.

How Does Hand-Object Detection Work?

The task consists of several sub-tasks:

• Detection of item pick from the shelf — hand crosses the shelf zone and grabs an object (item pick detection)
• Item identification — SKU-level recognition: what exactly was taken
• Shopper tracking — associating the picked item with a specific person
• Return detection — picked up → looked → put back

Each sub-task is non-trivial. Together, they require deep integration of CV and MLOps.

Hand-Object Detection Mechanism

The first step: find hands in the frame. MediaPipe Hands works well for isolated hands, but in retail, hands are often partially occluded by shelves, other items, or the shopper's body. A specialized hand detector — like 100DOH or a fine-tuned EgoHands dataset — performs better in cluttered scenes.

After hand detection, we recognize hand-object contact. This is not simply spatial bounding box overlap: we need to determine whether the hand is holding the object or just near it. Approaches:

• Bounding box overlap + velocity analysis: the object moves synchronously with the hand → contact
• Contact state classification: a separate classifier on pairs (hand crop, object crop) → holding/not holding
• Pose estimation: the palm is closed around the object — grasp detection via hand keypoints

What Are the Challenges in SKU-Level Recognition?

A store carries 5,000–50,000 unique SKUs. A classifier with 50,000 classes is unrealistic without special techniques. Our embedding-based recognition is 5x faster to add new items than traditional retraining of a closed classifier. Metric learning with ArcFace is 3x more accurate than softmax on large-scale SKU recognition.

An effective approach: Metric learning + open-set recognition. Instead of a closed classifier, we use embeddings where similar items are close in the space. Adding a new SKU means adding one reference embedding without retraining the model.

Backbone: ConvNeXt-Small or EfficientNet-B4 with ArcFace loss, trained on the product catalog (studio photos). Few-shot recognition: 1–5 reference photos per SKU are sufficient for reliable identification. On an internal benchmark (2,000 SKUs of a grocery store in real store conditions): top-1 accuracy 0.87, top-3 accuracy 0.96.

Additionally: barcode/QR code as a backup channel. If the item in hand is positioned conveniently, the CV system reads the code directly via ZXing or ZBar integrated into the pipeline. Confidence-based fusion: if barcode confidence > 0.98, we use it; otherwise, visual recognition.

Infrastructure for the Store

All cameras are time-synchronized with sub-millisecond accuracy — critical for tracking across cameras.

Comparison of Contact Detection Methods

Deliverables: What's Included in the Work

1. Store zone audit: layout, lighting, shelf types
2. Camera layout design considering blind spots
3. Development of hand-object detection and item pick detection pipeline (PyTorch, TensorRT)
4. Training an SKU-recognition model on your assortment
5. Integration with POS/scale systems (optional)
6. Documentation and staff training
7. 3 months of post-release support

Limitations and Realistic Expectations

Honest about difficulties:

• Errors with similar packaging: two variants of the same product (classic/light) — most challenging
• Partial visibility: the shopper blocks the item with their body — an inevitable gap
• Returns to the wrong shelf: we detected the pick, tracked, but the item was placed elsewhere

Accuracy at the level of Amazon Go (~99%+ per their claims) requires additional sensors — load cells on shelves remove uncertainty: N grams decreased → one specific pack taken. CV + scales significantly outperform pure CV.

The technology behind Amazon Go, described on Wikipedia, is based on computer vision and weight sensors.

Timeframes

Pilot system for a 20–30 m² zone with a limited assortment (200–500 SKUs): 10–16 weeks. Full system for a 100+ m² store with POS integration: 5–9 months.

We can assess your project: contact us for a free layout audit — we will send a technical proposal with estimated timelines. Request a project and get a consultation from our engineer. We will also help with selecting camera and server configurations for your budget. Contact us to discuss details.

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.