1 paper · 1 filter
Xiangshuai Song, Jun-Jie Huang, Tianrui Liu +2
Despite the success of convolution- and attention-based models in vision tasks, their rigid receptive fields and complex architectures limit their ability to model irregular spatia…