Computer vision-based DC warehouse imaging for automatic volume estimation

Overview & Synopsis

Matterport models enable a user to visualize and evaluate a space using panoramic views and building scenes. In our work we explore the use of the Matterport API to train a Mask R-CNN model on a warehouse images dataset. Thereby allowing us to compute dimensions and volume automatically in a computer vision model. Optimal space sizing can bring operational and cost benefits in supply chain logistics and planning. Our goal is to develop a trained model that can identify storage racks in a warehouse in images and video. Prior work in this area such as the COCO model have been trained to detect 80+ object categories. However, when the pre-trained models were used to image warehouses they were found to underperform, unable to identify objects in an image due to lack of training information. Carton pallets in the warehouse were identified incorrectly as couch, chair and suitcase. It is envisioned in the future that data gathering would use the Matterport RGB-D dataset such as Lloyd's Antiques Aoyama -Virtual Tour leading to richer trained models. There are no datasets in the market today for warehouses or Distribution center indoors imaging. We present a first to market, industry first use case. AR-VR gear can be used to capture and train a live image feed in a DC warehouse as well as to fine tune the model.