Automatic detection and classification of products on supermarket shelves

Nowadays, the business expansion of the retail industry is leading to an unprecedented rise in the amount of items which are stocked and sold in supermarkets. In such a context, the automatization of the product identification and categorization process, allow easier and faster management of the inventory, providing relevant information concerning the stored goods. To address this challenge, deep learning algorithms have emerged as a promising solution for accurately identifying and recognizing grocery items. This thesis work aims, therefore, to describe the development of a system trained for performing the automatic classification and detection of products on supermarket shelves. Due to restrictions related to our real world use-case and in order to take advantage of the versatility of such approach, we decided to implement an algorithm based on detection and classification carried out in two consecutive steps. Regarding the detection, we initially trained a CenterNet architecture on the public SKU-110k dataset, and then we applied a fine-tuning on our previously annotated private dataset. Concerning classification instead, we experimented with a wide range of different methods with the purpose of performing a comparison and choosing the most effective one. Lastly, we assembled the best detector-classifier combination in order to develop an End-to-end system that archived the 56.67% of mAP on our private benchmark.