Computer Vision News - January 2022

10 Best Paper Award BMVC DISCO: ACCURATE DISCRETE SCALE CONVOLUTIONS Ivan Sosnovik and Artem Moskalev are PhD students at the University of Amsterdam in the UvA Bosch Delta Lab, under the supervision of Professor Arnold Smeulders. Last month, they won the Best Paper award at the British Machine Vision Conference (BMVC) for DISCO, their innovative new approach to accurate scale-equivariant convolutional neural networks. They are here to tell us more about it. We know objects in images and video vary in size if the distance between the object and the camera changes. When watching a football game, for example, we assume the distance between us and the players increases or decreases as they move away from or towards the camera. “ Convolutional neural networks are currently the state-of- the-artmodels for respondingtothesesizedifferences, ” Ivan tells us. “ They are so good because of one very important feature: convolutions are equivariant to translations . In other words, a translation of the input signal leads to a proportional translation of the output signal. ” Ivan and Artem want to take this one step further by equipping a CNN with an extra quantity – scale . They want to achieve a model for which a translation and scale transformation of the input will lead to a proportional translation and scale transformation of the output to analyze images and all their rescaled copies in the same way. Other papers have tried different methods to build such networks, but DISCO presents a theory which is generalized and at the same time shows us there is a missing element in this field of CNNs for scale equivariance. This missing element was a set of kernels, a class of convolutions, which are very B E S T A W A R D PAPER Artem Moskalev Ivan Sosnovik