Increasing accuracy 90% for a big box shipper
The solution used was transferred learning harnessing CNN models. We trained with image semantic segmentation and make it more sensitive to packages in a 2d environment. Then training the model (3D Bounding Box Estimation) to give our model the human capability of 3D concept of 2D images. We created a diverse data set to train the model to predict the percentage of interior volume. Finally we moved the model to the ML (machine learning) kit to make it work locally and natively on the Android device using augmented reality.
Producing 90% accuracy which matched similar to expert measurements.