CVPR 2022 High quality paper sharing

A ConvNet for the 2020s

The paper ：https://arxiv.org/abs/2201.0354

Code ：https://github.com/facebookresearch/ConvNeXt

2020 Since then ,ViT It has always been a research hotspot .ViT The performance of image classification exceeds that of convolutional network , Various variants developed later will ViT Carry forward （ Such as Swin-T,CSwin-T etc. ）, It is worth mentioning that Swin-T The sliding window operation in is similar to the convolution operation , Reduced computational complexity , bring ViT It can be used as a backbone network for other visual tasks ,ViT It's getting hotter . This paper explores where the convolution network loses , Where is the limit of convolution network . In this paper , The author gradually moved towards ResNet Add structure （ Or use trick） To improve the performance of convolution model , In the end ImageNet top-1 It's done 87.8%. The author believes that the network structure proposed in this paper is a new generation （2020 years ） Convolution network of （ConvNeXt）, Therefore, the article is named “2020 Convolution network in the s ”.

Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding

The paper ：https://arxiv.org/abs/2203.00867

Code ：https://github.com/DQiaole/ZITS_inpainting

In recent years , Great progress has been made in image restoration . However , It is still challenging to restore damaged images with vivid texture and reasonable structure . Because of convolutional neural networks (CNN) My receptive field is limited , Some specific methods can only deal with regular textures , The overall structure will be lost . On the other hand , Attention based models can better learn the remote dependence of structural recovery , But they are limited by a large number of calculations in large image size reasoning . To solve these problems , This paper proposes to use additional structure restorer to gradually promote image restoration . The proposed model uses a powerful attention based method in a fixed low resolution sketch space Transformer Model to restore the overall image structure .

Class Re-Activation Maps for Weakly-Supervised Semantic Segmentation

The paper ：https://arxiv.org/pdf/2203.00962.pdf

Code ：https://github.com/zhaozhengChen/ReCAM

This paper introduces a very simple and efficient method ： Use the name ReCAM Of softmax Cross entropy loss (SCE) Reactivate with BCE Convergence of CAM. Given an image , This article USES the CAM Extract the characteristic pixels of each class , And use them with class tags SCE Learn another fully connected layer （ After the trunk ）. After convergence , This paper is based on CAM Extract... In the same way as in ReCAM. because SCE The contrasting nature of , Pixel responses are decomposed into different categories , Therefore, the expected mask ambiguity will be less . Yes PASCAL VOC and MS COCO Our assessment shows that ,ReCAM It can not only generate high-quality masks , It can also be in any CAM The variant supports plug and play with little overhead .