Based on input images from warehouses, the objective is to detect situations in which pallets and loads are potentially unsafe or abnormal. That means the algorithm needs to be able to recognize if there is a pallet, if there is a load, if the pallet is damaged, if the load is damaged, and if the image as a whole is abnormal.
This problem is difficult due to the presence of high variation in pallet type, packaging material, camera angle, occlusion, lighting, and background clutter. There could also be a possibility of more than one pallet in a single image, which makes reasoning about the entire image very difficult.
To address this, we formulate pallet anomaly detection as a structured visual reasoning problem. Instead of only detecting objects, the system must extract meaningful attributes from each detected region and aggregate them into an image-level safety decision.
