【论文汇总】人群计数中Transformer的应用，持续更新

1.CCTrans: Simplifying and Improving Crowd Counting with Transformer

论文地址：https://arxiv.org/pdf/2109.14483v1.pdf

代码：https://github.com/wfs123456/CCTrans
全监督+弱监督

Transformer backbone：Twins

我的解读：CCTrans: Simplifying and Improving Crowd Counting with Transformer解读和代码踩坑记录_算法卷死我算了的博客-CSDN博客

2.TransCrowd: Weakly-Supervised Crowd Counting with Transformer

论文：https://arxiv.org/abs/2104.09116

代码：GitHub - dk-liang/TransCrowd: TransCrowd: Weakly-Supervised Crowd Counting with Transformer

弱监督 ，Transformer backbone:VIT

我的解读：TransCrowd: Weakly-Supervised Crowd Counting with Transformer解读_算法卷死我算了的博客-CSDN博客

3.Joint CNN and Transformer Network via weakly supervised Learning for efficient crowd counting 

论文：https://arxiv.org/abs/2203.06388

弱监督，Transformer backbone:Swin transformer

4.An End-to-End Transformer Model for Crowd Localization

论文：https://arxiv.org/abs/2202.13065

人群定位的任务，decoder也用的transformer，无代码，Transformer backbone:应该是DETR

5.Congested Crowd Instance Localization with Dilated Convolutional Swin Transformer

人群定位的任务，encoder用的Swin transformer，decoder是金字塔卷积，无代码。 

论文：https://arxiv.org/pdf/2108.00584.pdf

6.Boosting Crowd Counting with Transformers  

论文：https://arxiv.org/abs/2105.10926

 写的一言难尽，没有代码，感觉是把这篇文章Tokens-to-token ViT: Training vision transformers from scratch on imagenet用在了crowd counting上面。

Tokens-to-token ViT中提到的deep narrow VIT和VIT的区别主要在于更深以及维度更小

7.CrowdFormer: Weakly-supervised Crowd counting with Improved Generalizability 

论文：https://arxiv.org/abs/2203.03768

弱监督。backbone是PVT-V2(PVT-V1的升级版)，回归只用了一个线性层，盲猜和TransCrowd一个回归，应该是就用了linear来做回归。

8.Boosting Crowd Counting via Multifaceted Attention（CVPR2022）

论文：https://arxiv.org/abs/2203.02636v1

作者解读：【可学区域注意力】Boosting Crowd Counting via Multifaceted Attention - 知乎

代码： 已公布GitHub - LoraLinH/Boosting-Crowd-Counting-via-Multifaceted-Attention: Official Implement of CVPR 2022 paper 'Boosting Crowd Counting via Multifaceted Attention'

9. Counting Varying Density Crowds Through Density Guided Adaptive Selection CNN and Transformer Estimation

论文：https://arxiv.org/abs/2206.10075

arxiv尚为公布代码，backboneVGG-16,CNN+Transformer的双分支

10.【CUT】Segmentation Assisted U-shaped Multi-scale Transformer for Crowd Counting

论文：发表于BMVChttps://www.researchgate.net/publication/364030579_Segmentation_Assisted_U-shaped_Multi-scale_Transformer_for_Crowd_Counting

Backbone是：Twins-PCPVT，思路和cctrans有点像，cctrans用的Twins里的Twins-SVT-L作为transformer backbone。loss没什么新意，拼了dm-count和Focal loss.不过对每个阶段的特征都进行监督。SAAM模块里面做了一个语义分割的额外任务，mask会放到简单transformer block里面学习。

 

11.MAFNet: A Multi-Attention Fusion Network for RGB-T Crowd Counting

论文：https://arxiv.org/abs/2208.06761

RGB-T（输入为RGB图像和热图）人群计数任务 。backbone为VGG19,在后面的模块中用了self-attention和跨模态的self-attention。

我的解读：【MAFNet】 A Multi-Attention Fusion Networkfor RGB-T Crowd Counting解读_行走的人参的博客-CSDN博客

12.Spatio-channel Attention Blocks for Cross-modal Crowd Counting（ACCV2022）

论文链接:

https://openaccess.thecvf.com/content/ACCV2022/papers/Zhang_Spatio-channel_Attention_Blocks_for_Cross-modal_Crowd_Counting_ACCV_2022_paper.pdf 代码：https://github.com/VCLLab/CSCA 

 跨模态的双分支网络，CSCA模块中采用了non-local的跨模态的注意力。和第11篇挺像。挺简单的，代码改编自RGBT-CC数据集那篇（CVPR2021）。

13. Attention-Guided Collaborative Counting

发表在：IEEE Transactions on Image Processing 

论文链接：Attention-Guided Collaborative Counting | IEEE Journals & Magazine | IEEE Xplore 

 提出了一个注意力引导模块和协同计数模块，双transformer（空间transformer和通道transformer）