A Fortune 500 shipping company built models to read truck fill levels from images. Labelbox produced the object-detection signal and shipped production models in five months.

The challenge

A Fortune 500 shipping and supply-chain company wanted to use machine learning to assess the state of packages inside its shipping vehicles. By reading truck fill levels with computer vision, it could maximize the volume it ships and optimize the trailer-unloading process. The bottleneck was signal: the data science team couldn't produce all the high-quality labeled data its computer vision models needed, had no scalable in-house labeling, and lacked software to visualize unstructured images in one place to prioritize what to label, find edge cases, and skip duplicates.

The approach

The company chose Labelbox to produce the signal and see what data it had. With Annotate and Catalog, it prioritized object detection and bounding-box projects — classifying truck-interior packages and fill rates and sorting by metadata like camera number, date/time, and unloading stage. With Labelbox Boost, the team preprocessed images by importing baseline predictions: it labeled a small set against a specific trailer-unloading process, then propagated those labels across hundreds of thousands of images. Production-ready models shipped in five months.

The outcome

The baseline-prediction workflow cut roughly 50% of time and cost and reduced workloads by thousands of hours. Next, the team is expanding to more computer vision — robotics for sorting and identifying packages, and automating more manual and auditing processes.

Where this goes

Logistics is becoming embodied AI — robots that sort and load. The signal that teaches a model to read a truck interior is the same signal that will run the robots doing the work.