Investigation of Deep Learning Datasets for Warehousing Logistics

Deep Learning for Computer Vision holds great potential in warehousing logistics, for example for applications such as mobile robots or autonomous forklifts. However, the availability of labelled image datasets within this area is limited. To address this problem, we benchmarked two different datasets, LOCO (Logistics Objects in Context) and TOMIE (Tracking Of Multiple Industrial Entities), to find out, if these datasets can be used interchangeably. Therefore, we examine the usability of these datasets for Object Detection tasks using the YOLOv7 framework. For this we trained several networks and compared them with each other. A deep analysis between these two datasets shows that they are quite different and only suitable for specific tasks which are not interchangeable, despite having emerged from the same research domain. More thorough investigations are performed to find the reasons for this lack of compatibility. To close the gap between LOCO and TOMIE, a synthetic data generation pipeline for pallets is developed and 18,000 synthetic pallet images are rendered. Furthermore, models are trained based on the synthetic data and compared with the models trained on real data. The synthetic data generation pipeline successfully closes the reality gap, and the performance on TOMIE is increased, but the performance on LOCO remains significantly weaker, in comparison. To develop a deeper understanding of this behaviour we examine the underlying datasets and the reasons for the performance difference are identified.