Generating embedding spaces for re-identifying pallets from their chipwood patterns

Modern logistics rely on continuous tracking of even most basic entities like pallets. While tracking is nowadays achieved via additional markers (e.g., barcodes), pallets have a unique woodchip pattern in their blocks that can be exploited for re-identification if suitable methods are employed. This paper shows how this task can be accomplished in an open-set scenario. In this work, images of pallet blocks are transformed into vector representations using neural networks based on the ResNet-50 architecture. The networks are trained such that the Euclidean distance between image vectors of the same palette block is small. Several approaches, i.e., cross entropy, contrastive and triplet loss, to create the embedding space are compared. It is shown that Triplet Loss has advantages over other methods with a top-1 accuracy of 0.86 on the pallet-block-32965 dataset. In addition, a neighborhood-based approach to novelty detection called NSAND is presented, which outperforms a purely distance-based approach by up to 20%.