[Paper Abstract] screenshots of methods for recording abstracts of interest and papers in special fields.

Target segmentation

1. ISDNet: Integrating Shallow and Deep Networks for
   Efficient Ultra-high Resolution Segmentation【CVPR2022】
   
   The huge burden of computation and memory is the two obstacles of ultra-high resolution image segmentation . To solve these problems , Most of the previous work followed the overall situation - Local refining pipeline , The pipeline pays more attention to memory consumption and ignores the reasoning speed . Compared with the pipeline that divides the large image into small local areas , We focus on directly inferring the entire image . In this paper , We proposed ISDNet, This is a novel ultra-high resolution segmentation framework , It integrates shallow and deep networks in a new way , While realizing accurate segmentation, it significantly accelerates the reasoning speed . In order to further utilize the relationship between shallow and deep features , We propose a novel relationship aware feature fusion module , It ensures the high performance and robustness of our framework . stay Deepglobe、Inria Aerial and Cityscapes A lot of experiments on data sets show that , Our performance is always better than the most advanced technology . say concretely , It's in Deepglobe In order to 27.70 FPS At a speed of 73.30 mIoU, More accurate than recent competitors , Fast 172 times .