Computer Vision News - April 2018

Computer Vision News Application 20 Application: Aquifi Aquifi develops scalable 3D computer vision solutions for application in manufacturing and logistics. In particular, they combine computer vision and deep learning techniques to speed up and improve the accuracy of visual inspection processes. They work across the Americas – North and South – and are opening up in Europe and Asia, in particular Japan and China. Carlo Dal Mutto , CTO at Aquifi, tells us that the company’s focus shifted to manufacturing and logistics in 2016. They started off developing their depth camera and were originally targeting the consumer market through mobiles phones and tablets, but soon realized its applications like object sizing and 3D modelling were a good fit for e-commerce. At that point they began to talk to several e-commerce companies and realized there were a lot of problems they could help to solve. This opened them up to logistics, which is highly affected by the push on e-commerce and the rising need for shipment services. From there, they went into manufacturing and the more they were digging, the more problems they found to solve. From a logistics point of view, inspection of boxes and object sizing is important - sizing objects to find the right box and so that they can fit in the right truck for shipment. For manufacturing, it is more about visual inspection for identification and anomaly detection. In particular, if an object is on a mixed conveyor belt, it has to be identified and the different parts have to be identified to ensure they are all correct. Then each of those parts needs to be inspected for defects and anomalies. If there are anomalies, then the objects are selected and there are two ways to find them. One is if there is an operator at the end of the line, a GUI shows the operator that there is a defect and where it is. The other possibility is that they trigger a signal so that the object gets moved off the conveyor belt. Aquifi uses a variety of computer vision techniques , in particular, depth estimation 3D reconstruction and object recognition from 3D models . The depth estimation is achieved with an in-house designed active stereo system. They designed a pattern projector and have a couple of IR imagers synchronized with a color camera as well. Depth estimation from active stereo is a technique that they use a lot and is very reliable, both in terms of precision and accuracy across a variety of different materials. In the past, before the deep learning revolution, they used passive stereo techniques. “What we can do always is to incrementally build on top of what we currently have.”