Computer Vision News - January 2022

11 DISCO: accurate Discrete Scale Convolutions accurate and at the same time very flexible in the sense that you can choose any set of scales and it will be the most accurate model. “ DISCO uses a different type of kernels to other models, ” Ivan explains. “ These kernels are sparse. A big fraction of elements in these kernels are zeros. With PyTorch or similar frameworks, where there are no sparse convolutions, it’s not possible to implement these kernels to get the best out of the theory. We demonstrated the computational complexity of the network grows linearly with the increasing number of scales, but that’s not true for PyTorch because it doesn’t calculate sparse convolutions. ” One of the practical applications for this work is in autonomous vehicles . If an autopilot can identify every car model on the market, it must also be accurate at varying their size. It cannot be confused because a car is 500 meters away. This method is highly accurate at classifying and tracking objects, so in this example would be able to follow the cars highly accurately and significantly faster than all previous models with the same features. The system could also use those cars to deduce the geometry of the road the vehicle is driving on, which may not be the main focus of the autopilot but supports many auxiliary tasks. “ In this work, we designed a scale-equivariant network , and we show lower equivariant error leads to better accuracy, which means because we add additional structure to our neural network, it’s able to generalize better, ” Artem explains. “ It needs less data to generalize, so it’s useful in many areas where the transformation of the data is not limited to the translation group. ” Ivan is no stranger to awards. We last spoke to him in 2017 when he had just won the Kaggle Leaf Classification Competition . But it takes an extra special paper to take home a Best Paper award. What made DISCO stand out to the judges? “ That’s a very good question and one I’ve definitely thought about because I would like to repeat this success! ” he laughs. “ I think it was more about the set of experiments we did. We demonstrated