Keyword [DRN]
Yu F, Koltun V, Funkhouser T. Dilated residual networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 472-480.
1. Overview
1.1. Motivation
- progressively reduce resolution is harmful for segmentation, even classification
In this paper, it proposed DRN (Dilated Residual Network)
- increase the resolution of output feature maps without reducing the receptive filed of individual neurons
- removing gridding artifacts introduce by dilation
2. Dilated Residual Networks
- remove the stride-2 in layer-4 and layer-5
- set dilation = 2 of block_i in layer-4, i≥2
set dilation = 2 of block_1 in layer-5
set dilation = 4 of block_i in layer-5, i≥2
3. Degridding
- removing max pooling
- adding layers
- removing residual connections
4. Experiments
