PAD-Nets: Learning Dynamic Receptive Fields via Pixel-Wise Adaptive Dilation

Dongdong Wang, Hao Hu, Jie Yao, Zihang Zou, Liqiang Wang

Keywords: receptive field, dilated CNN, representation learning

Abstract: Dilated convolution kernels are constrained by their shared dilation, keeping them from being aware of diverse spatial contents at different locations. We address such limitations by formulating the dilation as trainable weights respect to individual positions. We introduce Pixel-wise Adaptive Dilation (PAD), a light-weighted extension that allows convolution kernels to flexibly adjust receptive fields based on different contents at pixel level. By inferring dilation via modeling inter-layer patterns, PAD-Nets also provide a possible way to partially understand the hierarchical representations of CNNs. Our evaluation results indicate PAD-Nets can consistently outperform their conventional counterparts on various visual tasks.