How to Setup and Maintain Image Recognition

in order for the system to correctly recognize data from a photo, it needs initial information about the products the eyrene performs image recognition quality (qa) with periodic neural net retraining usually, 5 10 times a month depending on data change rates; automated image recognition accuracy evaluation using labeled tests sets; automated image recognition accuracy evaluation based on feedback from mobile users; semi manual image accuracy evaluation this is based on image recognition internal score metrics, feedback from mobile users, and automatic measurements; semi manual corrective measures such as new product design variation introduction, new competitor brand introduction, and additional data search/labeling/cleaning the image recognition accuracy evaluation report is provided by the eyrene on weekly basis how it works for the initial setup of the system, you will need to provide eyrene staff with information about your products and those of competitors this is necessary so that the system can learn to correctly recognize products, their parameters, tastes, sizes, etc sequencediagram participant client participant eyrene client >>+eyrene 1 initial setup data note right of eyrene 2 training image recognition eyrene >> client new product catalog client >>+eyrene 3 new products or product designs note right of eyrene 4 re training image recognition eyrene >> client updated catalog on the portal initial installation data provide eyrene employees with information about your products and their hierarchy, equipment, posm, etc training image recognition after you provide the information, employees will begin training the system to recognize photos with maximum accuracy the training period takes 2 4 weeks, depending on the initial completeness of the catalog at the same time, a new product catalog will appear on the portal new products or product designs when new products appear in the assortment or the design of products changes, you need to update the catalog re training image recognition after the data is sent to eyrene, employees will begin retraining recognition based on the new data it takes 5 working days at the same time, the product catalog on the portal will be updated initial setup for the initial setup of the system in a new market, provide eyrene employees hierarchy of your product and brand product catalog equipment catalog posm catalog competitor catalog shelf image archive hierarchy of own product and brand this is the number of hierarchical levels and complete hierarchies used it is usually provided in excel files product catalog this is the own trade items catalog with all relevant product attributes the product catalog can be imported from excel tables or received from api a minimum list of attributes primary identifier of ean/gtin or another identification system; primary name – in english; category — lowest level of product hierarchy used; brand — lowest level of brand hierarchy used; manufacturer ; ean/gtin or ean list country list in iso alpha 2 ; product physical dimensions in mm height, width, depth; at least one catalog image several images can be provided to show the product from different sides or if the product has design variations catalog product images can be both real photos or marketing images image height must be larger than 400 px, high resolution is not required additionally, you can provide product attributes relevant to image recognition — such as product package type or size (volume/weight/etc); product attributes relevant for kpi calculations/aggregation/analysis; a list of local product names — to localize mobile app messages to the agents equipment catalog this is the own equipment catalog at the level, relevant for image recognition note you don't need to specify each cooler model if the recognition is supposed to recognize 1 and 2 door coolers, the same for other equipment a minimum list of attributes • primary code ; • primary name – in english; • equipment type (depending on equipment classification used); • at least one catalog image the requirements are the same as for the product catalog posm catalog own posm catalog at the level of detail, relevant for image recognition, data is similar to equipment catalog data competitor catalog catalog with competitor brands for each relevant category (lowest level of product hierarchy) competitors can be provided brand + category level or sku level if required for kpi calculations for example, the “price pair index” kpi measures own product price/competitor product price the vendor could provide their existing competitor catalog depending on the market and category a minimum list of attributes category — lowest level of product hierarchy used; brand ; logo — at least one image or link to the image image height must be larger than 400 px in the process of initial setup and system operation, the eyrene will extend the competitor catalog with the most common unknown brands seen in the shelf images in the course of the image recognition qa process note unknown competitor products can still be recognized at the category level (e g “unknown beverage”), though the accuracy of such recognition is different for different categories and different product hierarchy levels shelf image archive archive with 1,000 3,000 images per macro category per market the recognition is trained using real shelf photos you will not need to label the images on the shelves the system will do it automatically shelf images must be at least 2 mpx in resolution, must be relatively sharp, and do not have large perspective distortions checklist hierarchy of own product and brand excel file with product and brand hierarchy product catalog attributes primary identifier in ean/gtin primary name in english manufacturer ean/gtin or ean list country list in iso alpha 2 product physical dimensions in mm (h/w/d) at least one catalog image images for products with different design variations, a photo of each of the variations has been added the height of each image is over 400px optional added additional product attributes type, product packaging size, weight, volume, etc added additional attributes for kpi added list of local product names equipment catalog attributes primary code primary name in english equipment type at least one catalog image images for equipment with different design variations, a photo of each of the variations has been added the height of each image is over 400px posm catalog attributes primary code primary name in english posm type at least one catalog image images for posm with different design variations, a photo of each of the variations has been added the height of each image is over 400px competitor catalog attributes category with lowest level of product hierarchy used brand logo — at least one image or link to the image images the height of each image is over 400px optional competitors provided brand + category level or sku level shelf image archive the archive contains a minimum of 1000 images per macro category for each market each shelf image in the archive have at least 2 mpx resolution maintenance catalog updates to keep the product catalog in eyrene up to date, you need to establish a process of periodic updating for example, regular weekly updates plus infrequent urgent unscheduled catalog updates if new products appear or product design changes, the catalogs also need to be updated catalog updates can be provided as an excel file + image or as an api endpoint alternatively, you can update the catalog using our api endpoint or manually in the portal raci matrix activity description values initial setup data set of objects and parameters required, e g quantity/quality/types of pictures, list of physical characteristics, etc the information is to be provided for the objects that can be recognized by the vendor’s solution, incl customer trade items (and other product dimensions like brands) competitor products customer equipment and posm the vendor should clearly specify which kind of pictures the network gets initially trained (and re trained in the operation mode) ideal images of the product with 360o turnaround used usually real images of the product in a store placement see above initial setup duration training period to achieve the maximum accuracy 2 4 weeks depending on initial catalog completeness period for change re training period to achieve the maximum accuracy 5 business days for a batch of new products/product designs accuracy measurement the vendor should specify how the accuracy of the solution gets measured (automatically, semi manually) in the operation mode and if there are any efforts expected from the customer side to execute that see above, no effort is expected from your side except of regular catalog updates