Emos Yalp

https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136780420.pdfAbstractMost image denoising networks apply a single set of static convolutional kernels across the entire input image. This is sub-optimal for natural images, as they often consist of heterogeneous visual patterns. Dynamic convolution tries to address this issue by using per-pixel convolution kernels, but this greatly increas..

https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136810034.pdfAbstractRecently, transformer and multi-layer perceptron (MLP) architectures have achieved impressive results on various vision tasks. However, how to effectively combine those operators to form high-performance hybrid visual architectures still remains a challenge. In this work, we study the learnable combination of convoluti..