Semi-automated data annotation for parcel detection in warehouses with Zero-Shot image segmentation methods: an evaluation

In face of a shrinking workforce, logistics companies need to automate low-skill high-effort tasks in order to remain competitive in terms of market share and job attractiveness. This often includes camera-based detection of parcels. While recent advances in AI-based image processing enable the training of powerful object detection algorithms, acquiring the required annotated data set is tedious and cost-intensive. We test the ability of Segment-Anything type zero-shot segmentation models in assisting data annotation for parcel detection. Evaluation on the publically available Stacked Carton Dataset shows that over 93% of segmentation masks can be extracted successfully with these models while also providing insights for future improvements.