search

ydhongHIT / DDRNet

The official implementation of "Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes"
MIT License
415 stars 53 forks source link

DAPPM + DDRNet slim for small image segmentation input_size (e.g. 128x128 or 256x256) #21

Closed tamirDia closed 2 years ago

tamirDia commented 2 years ago

What should be the kernel sizes+strides for DAPPM module for small image sizes? and the number of spp planes? Thanks :)

ydhongHIT commented 2 years ago

What should be the kernel sizes+strides for DAPPM module for small image sizes? and the number of spp planes? Thanks :)

In the paper, the stride is 2^n and the kernel size is 2^(n+1) + 1. Because the lowest resolution is 1x1, the resolution of feature maps passed into dappm determines the number of branches. In fact, the advantage of dappm is more obvious when the input image is larger. If the image resolution is 256x256, the input resolution of dappm is 8x8(remove the downsample of the last block). In this case, the kernel sizes and strides are 5, 2; 9, 4, and the last is global pooling.