AI Powered Packing Verification
Automate item detection for E-Commerce consignments to ensure complete, accurate deliveries.
The Opportunity
A leading e-commerce client experienced recurring issues with deliveries containing missing items, impacting supply chain integrity and customer satisfaction. Traditional approaches to packing verification relied on manual checks or generic image analysis, often failing to accurately flag consignments with missing goods before dispatch. The client sought a robust automated solution that could identify and flag any deliveries with missing items, given images of a tote on a conveyor before the goods are packed.
What we did
✔Advanced Model Selection: Evaluated off-the-shelf computer vision APIs, such as Google Vision API Product Search, and self-supervised models like DINOv3 to detect both visible and obscured items using fine-grained segmentation. ✔Image Pre-processing: Cropped images to remove barcodes and box edges to reduce false positives, and applied augmentations (e.g. rotation) to improve detection coverage and accuracy. ✔Post-Processing Enhancement: Applied colour clustering and confidence thresholds to filter out low-confidence or irrelevant detections, ensuring precise and reliable item counts. ✔OCR & Brand Detection Integration: Used OCR and logo detection to cross-verify item counts, identifying text clusters and brand marks for added validation. ✔Iterative Improvement: Continuously fine-tuned model parameters and expanded the image library to enhance precision and reliability across complex item arrangements.
The Results
Reduced Missing Item Incidents: Consignment accuracy was markedly improved, with both uniform packaged goods and previously hard-to-detect, obscured items reliably identified. Operational Efficiency: Automation of the verification process reduced manual intervention and associated labour costs. Enhanced Customer Trust: Improved fulfilment accuracy led to higher customer satisfaction and fewer complaints relating to missing items.
>13
weeks to deliver a fully operational AI-powered item detection system.
How we did it
The system was developed using Python, leveraging PyTorch for model deployment and training, with a particular focus on self-supervised segmentation models such as DINOv3. These models enabled accurate detection of both visible and partially obscured items, ensuring robust performance across varied packing scenarios. To enhance detection capabilities, cloud-based computer vision APIs—including Google Vision—were integrated to support object detection and OCR functionality. This combination of in-house and cloud-based components provided flexibility, scalability, and high accuracy across multiple product types. Image preprocessing workflows were designed using OpenCV, incorporating cropping and augmentation techniques to optimise detection performance. Additional clustering and post-processing routines were implemented to filter detections and maintain precise, reliable results. The entire pipeline was engineered for seamless integration into the client’s existing packing infrastructure. By leveraging cloud resources and containerised deployment, the solution ensured scalable, efficient, and easily maintainable processing, allowing the retailer to automate verification at scale with minimal disruption to existing operations.
