Tag: Machine Learning

By Philipp Schmid, CSEM, Andrea Dunbar, CSEM and Jakob Olbrich, PwC

Meeting of Expert Group Machine Learning Clinic, February 11 2022

What have expensive mechanical watches, sand, e-waste and cockpits in common? All areas have tough challenges in computer vision. Human eyes are very hard to outperform with cameras and image processing. What people perform with their visual sense every day is just amazing and and creating these capabilities remains a complex challenge for computer vision.

At this first in person meeting this year the expert group focused on various real world vision problems.
The event was hosted by PwC in their inspiring location in Oerlikon. Four speakers set the floor for great discussions followed by a lively sitting Apéro.

Lukas Schaupp, PwC «Detecting e-Waste»
The amount of electronic devices people dispose is growing exponentially. Not just talking about smartphones, laptops and earphones but as well larger household items like dishwashers, toasters and vacuum cleaners. As prices for raw materials are rocketing off automated recycling of e-waste is becoming attractive. Lukas demonstrated strategies to localize and classify different electronic devices in bulk on a conveyor belt.

Andrea Dunbar, CSEM «AI at the Edge – Safety in the next generation Cockpits»
There are multiple reasons and advantages to process at the edge. Andrea demonstrated this impressively in the use-case: next generation cockpits. Pilot drowsiness detection and more important high accuracy eye gaze detection (±1°) with rates of up to 60 frames per second are only possible at the edge. What is today already reality in the flight simulator will soon be introduced in each car for the safety of our roads.

Francesco Cicala, PwC «Automatic image thresholding for semantic segmentation»
The quality of concrete depends heavily on the right mixture of sand and pebbles. In the future a smartphone app should be able to classify the correct mix by assessing the size of the sand and pebbles. Francesco introduced a powerful method to extend Otsu’s thresholding technique into a locally adaptive threshold map for the whole image. This method is robust, fully explainable and there are no labels needed. In a next phase it will be extended with a U-Net algorithm to improve accuracy.

David Honzatko, CSEM «Photometric stereo in defect detection»
Swiss Made symbolizes perfect quality. Especially in the watch industry requirements are demanding. The small parts are highly reflective, complex shaped and defects can appear randomly at any position. The key to an automated defect detection solution is photometric stereo. David presented a dome setup which can project up to 108 illumination directions. To reduce the hardware requirements whilst keeping the performance David presented a new data augmentation technique, which boosts the training of any deep learning architecture processing the images.

A full evening of new insights and tough challenges in the field of computer vision. Thanks to everyone
for the great participation and especially to the host for the amazing location and the local Apéro.