DISCO: accurate Discrete Scale Convolutions
This work addresses the challenge of scale as a disturbing factor in vision tasks, offering improvements for applications like classification and tracking, though it is incremental by building on existing scale-equivariant methods.
The paper tackled the problem of achieving accurate scale equivariance in convolutional neural networks for vision tasks requiring high granularity and small kernels, by deriving constraints for discrete scale convolutions and providing exact or approximate solutions, resulting in new state-of-the-art classification on MNIST-scale and STL-10 datasets.
Scale is often seen as a given, disturbing factor in many vision tasks. When doing so it is one of the factors why we need more data during learning. In recent work scale equivariance was added to convolutional neural networks. It was shown to be effective for a range of tasks. We aim for accurate scale-equivariant convolutional neural networks (SE-CNNs) applicable for problems where high granularity of scale and small kernel sizes are required. Current SE-CNNs rely on weight sharing and kernel rescaling, the latter of which is accurate for integer scales only. To reach accurate scale equivariance, we derive general constraints under which scale-convolution remains equivariant to discrete rescaling. We find the exact solution for all cases where it exists, and compute the approximation for the rest. The discrete scale-convolution pays off, as demonstrated in a new state-of-the-art classification on MNIST-scale and on STL-10 in the supervised learning setting. With the same SE scheme, we also improve the computational effort of a scale-equivariant Siamese tracker on OTB-13.