SAMDConv: Spatially Adaptive Multi-scale Dilated Convolution
PATTERN RECOGNITION AND COMPUTER VISION, PRCV 2023, PT VIII(2024)
摘要
Dilated convolutions have received a widespread attention in popular segmentation networks owing to the ability to enlarge the receptive field without introducing additional parameters. However, it is unsatisfactory to process multi-scale objects from different spatial positions in an image only by using multiple fixed dilation rates based on the structure of multiple parallel branches. In this work, a novel spatially-adaptive multi-scale dilated convolution (SAMDConv) is proposed to adaptively adjust the size of the receptive field for different scale targets. Specifically, a Spatial-Separable Attention (SSA) module is firstly proposed to personally select a reasonable combination of sampling scales for each spatial location. Then a recombination module is proposed to combine the output features of the four dilated convolution branches according to the attention maps generated by SSA. Finally, a series of experiments are conducted to verify the effectiveness of the proposed method based on various segmentation networks on various datasets, such as Cityscapes, ADE20K and Pascal VOC. The results show that the proposed SAMDConv can obtain competitive performance compared with normal dilated convolutions and depformable convolutions, and can effectively improve the ability to extract multi-scale information by adaptively regulating the dilation rate.
更多查看译文
关键词
SAMDConv,Image Segmentation,Receptive Field
AI 理解论文
溯源树
样例
![](https://originalfileserver.aminer.cn/sys/aminer/pubs/mrt_preview.jpeg)
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要