Belgoprism’s manual and labor-intensive stock management undergoes a transformation with Faktion’s AI-driven approach, integrating computer vision, weight measurement, and advanced color matching. This article explores the impact of these AI components on their inventory process.
Introduction
In the vibrant and growing industry of adult toys, Belgoprism has established itself as a notable player. With an array of 4,000 to 5,000 different toys made from liquid PVC and silicone, this company has experienced consistent growth, particularly in its transition towards the B2C market. This growth journey demands a radical shift from traditional methods to more sophisticated, tech-driven solutions.
Challenges and Complications
As Belgoprism expanded, so did the complexity of their operations. Managing a large and diverse inventory, the company relied on traditional, manual methods for stock management. Each toy required individual inspection and sorting based on size, color, and model – a time-consuming task that grew more daunting with each new product introduction.
The manual stock management system was not only time-consuming and resource-intensive but also prone to errors. The need for two to five full-time expert operators for quality inspection, manual sorting, and stock classification highlighted a process that was inefficient and unsustainable, making it clear that a more streamlined and accurate system was needed.
Identifying the Need
Recognizing these challenges, Belgoprism sought an automated solution to enhance their stock management process. The ideal solution would not only automate and optimize their processes but also complement their skilled operators, allowing them to focus on value-added tasks instead of repetitive, manual work.
Faktion's AI-driven Solution
Faktion introduces a tailored AI solution designed to meet Belgoprism's specific needs. The system integrates three core components: a computer vision-based classification model for shape detection, a precise weighing scale for size confirmation, and an unsupervised similarity search model for color identification. All these components need to be combined in a single, on-premise system. This system combines hardware components (a weight sensor, camera’s, and a tablet) with a software component (a user interface on a tablet to validate and correct the AI model’s suggestions). This integrated approach aims to enhance the efficiency and accuracy of Belgoprism's stock management, streamlining their inventory process.
The Approach
The project commenced with a practical Proof of Concept (PoC). As taking images of all toys is unpractical, and 3D designs were available for all toys, Faktion set up a pipeline to generate synthetic images based on the 3D files that can be used for modeling. For every toy, an unlimited number of images could be generated by rotating the toys, thereby creating variance in the dataset. A demo was created where a user could upload a synthetic image for classification. The model classified the image and the interface showed an example of the predicted class from the training set, together with the prediction probability per class.​ This initial stage was crucial for demonstrating the feasibility of the AI solution and laid the foundation for further refinement based on real-world feedback and performance analysis.
Initial results and next steps
At the conclusion of the first phase, the outcomes of Faktion's AI model were promising, demonstrating a remarkable level of precision when analyzing unseen synthetic images. This high accuracy with synthetic inputs highlighted the model's adeptness at navigating the controlled conditions these images present. However, a notable decline in accuracy was observed when the model encountered unseen real-world images. This discrepancy underscores the inherent challenges of transitioning from synthetic to real-world data, which often contains a broader variety of variables and less predictability.
To address these challenges and bridge the gap between synthetic and real-world performance, the subsequent steps are focused on enhancing the model's robustness. This will be achieved by diversifying the training set with a more comprehensive collection of real images, thereby exposing the model to the complexity and variability of real-world conditions. Moreover, by incorporating additional context such as toy weight into the model's considerations, the aim is to narrow down the range of possible correct answers, further refining the model's accuracy and applicability to real-world scenarios. This approach not only seeks to improve the model's performance but also illustrates the critical importance of integrating diverse, real-world data to ensure AI models can effectively translate their capabilities to practical applications.
