Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 5, No. 2, 2023 

 

35 

A	Survey	of	Crowd	Counting	Algorithm	Based	on	
Domain	Adaptation	
Yanan Wang1, Fen Luo1, * 

1 School of Software, Henan Polytechnic University, China 
* Corresponding author: Fen Luo (Email: luofenjsj@hpu.edu.cn) 

 

Abstract: Crowd counting, the task of estimating the number of individuals in a crowded scene, has gained increasing attention 
in computer vision research. However, crowd counting remains a challenging problem due to the complex and diverse nature of 
crowd scenes. In recent years, domain adaptation has emerged as a promising approach to improve crowd counting performance 
by adapting a pre-trained model to a target domain with different characteristics. This paper provides a survey of domain 
adaptation-based crowd counting algorithms, including their methods, datasets, and evaluation metrics. Overall, domain 
adaptation shows great potential in improving the accuracy and robustness of crowd counting algorithms, and further research 
in this direction is expected to lead to more effective and practical crowd counting solutions. 

Keywords: Crowd counting, Domain adaptation. 
 

1. Introduction 
With the accelerating urbanization and continuous 

population growth, the problem of crowd counting has 
become a widely researched area. Traditional crowd counting 
methods mainly rely on manually designed features and 
models, which often fail to adapt to the changes and 
complexities of different scenarios, resulting in low counting 
accuracy. Therefore, crowd counting methods based on deep 
learning have been widely applied, achieving excellent results 
by learning more effective feature representations and 
building more accurate models. 

However, in practical applications, factors such as 
differences between different scenarios, distribution 
differences between different datasets, etc., may lead to a 
decrease in model performance, which requires domain 
adaptation. Domain adaptation is a method of transferring 
knowledge learned in the source domain to the target domain, 
improving the generalization performance of the model by 
solving domain differences. 

Therefore, this paper mainly reviews the research status 
and development trends of crowd counting algorithms based 
on domain adaptation. This paper categorizes domain-
adaptive crowd counting methods into two types: one based 
on distribution strategy framework, and the other based on 
image transformation strategy framework. Firstly, the two 
methods are briefly introduced, then the commonly used 
datasets and evaluation indicators in population counting are 
introduced, and finally, the development direction of 
population counting based on domain adaptation is 
reasonably predicted. 

2. Methods 

2.1. Methods based on the distribution strategy 
 The first type of distribution strategy-based framework 

mainly uses discriminators to distinguish the density maps 
generated from the source and target domain images to align 
the data distribution.  

Wang et al. [1] proposed an adversarial learning method, 
namely CODA (adaptation to counting objects through scale-

perceived adversarial density). To handle different target 
scales and density distributions, through adversarial training 
by adapting multi-scale pyramid pyramid patches from the 
source and target domain, consistent count objects can be 
produced for different scales along with the ranking 
constraints of the pyramid input level.  

Yan et al. [2] solves the problem of domain shift by training 
using synthetic images and their associated labels and 
unlabeled real images, so that it can be born as a label for the 
purpose of fine-tuning the network.  

Zou et al. [3] proposed an adversarial scoring network 
(ASNet) that gradually closes the cross-domain gap from 
coarse-to fine-grained. The coarse-grained phase designs a 
dual discriminator strategy to adapt the source domain from 
the global and local feature space perspectives close to the 
target domain. In the fine-grained stage, the transferability of 
source features is explored by scoring the similarity of source 
and target samples from multiple levels based on the 
generation probability derived from the coarse-grained stage. 

2.2. Deep learning methods 
The second type of image transformation strategy-based 

framework is mainly based on Cycle GAN[4], converting the 
synthesized images into images with a similar style to the 
target domain, and then training the transformed images in the 
population counting task.  

Wang et al. [5] developed a data collector and hashtag that 
can generate synthetic population scenarios and annotate 
them without the need for manual labor. Based on this work, 
a large-scale, diverse synthetic dataset is constructed.  

Gao et al. [6] proposed a DACC (Domain Adaptive Crowd 
Counting) framework composed of high-quality image 
transformation and density map reconstruction. The former 
transforms synthetic data into real images to improve 
conversion quality by separating domain sharing/ 
independent features and design content perceived 
consistency loss, while the latter aims to generate pseudo-
labels on real scenes, using labels to retrain the final counter 
to improve count performance. 



 

36 

3. Datasets 
ShanghaiTech[7]: The dataset consists of two subsets: 

PartA and PartB. PartA contains 482 images with different 
resolutions, with 300 in the training set and 182 in the test set. 
The numbers in each image ranged from 33 to 3139, with an 
average of 501. The PartA images contain a variety of scenes, 
including indoor and outdoor scenes, and are mostly crowded 
images. Similarly, PartB contains 716 images with a 
resolution size of 768 * 1024, with 400 in the training set and 
316 in the test set. The number of people per image ranged 

from 9 to 578, with an average of 123 people per image. The 
images of PartB are taken from the city streets of Shanghai.  

UCSD[8]: The dataset was collected from a surveillance 
video recording on campus. Contains 2,000 frames with a 
resolution of 158 * 238. Number per frame ranged from 11 to 
46, with an average of 25 people per frame. This dataset 
provides a region of interest ROI (Region of Interest, ROI) 
for each frame that ignores background interference, is 
trained using the # 601 to # 1400 frames, and the rest are 
tested. Figure 1 shows the population image and the ROI 
region of the UCSD. 

 

（a）UCSD dataset image （b）ROI regions provided by the dataset
 

Figure 1. Population image and the ROI region of the UCSD. 
 

Mall[9]: Similar to UCSD, the dataset was collected in a 
surveillance camera in a shopping mall. Mall contains 2000 
frame files with a resolution of 240 * 320. Each frame 
contained numbers ranging from 13 to 53, with an average of 
31 per frame. The dataset provides a ROI ignoring 
background interference for the frame files. Unlike the 
grayscale frames of the UCSD, the Mall frame file is colored 
and contains different changes in light conditions. In addition, 
Mall has perspective distortion and occlusion phenomenon. 

QNRF[11]: The dataset is composed of 1,535 images with 
about 1.25 million annotations. It ranged from 49 to 12865, 
with an average count of 815.4. Specifically, the training set 
consisted of 1,201 images, and the test set consisted of 334 
images. The images of this dataset species contain a wider 
range of scenarios and contain the most diverse viewpoint, 
density, and illumination changes. 

4. Evaluating Indicator 
Mean absolute error (MAE) and root mean square error 

(RMSE) are two common indicators of population counting 
algorithm. MAE reflects the accuracy of model prediction, 
defined as formula (1), RMSE reflects the robustness of 
algorithm prediction and defined as formula (2): promising 
results in crowd counting, but further research is needed to 
develop more efficient and effective techniques for this 
challenging problem. 

 

1

1
| | #

N
GT

i i
i

MAE C C
N 

  （1） 

 

2
i

1

1
( ) #

N
GT

i
i

RMSE C C
N 

  （2）

 

5. Conclusion 
Although domain adaptation-based crowd counting 

algorithms have achieved good results, there are still many 
problems. Firstly, most methods are limited to single-source 
domain adaptation. However, in practical scenarios, domain 
shift not only exists between source and target domains, but 
also between different source domains. Data from different 
source domains can interfere with each other during mutual 
learning. Therefore, it is necessary to solve the problem of 
multi-source domain adaptation in crowd counting. Secondly, 
current domain adaptation methods are tested on crowd 
counting datasets in normal environments. However, there 
may be adverse environments in real-world scenarios, such as 
low-light, rain, and fog, and the model cannot effectively 
handle these scenarios. Therefore, the generalization ability of 
the model in adverse environments needs to be further 
improved. 

References 
[1] Li W, Yongbo L, Xiangyang X. Coda: Counting objects via 

scale-aware adversarial density adaption[C]//2019 IEEE 
International Conference on Multimedia and Expo (ICME). 
IEEE, 2019: 193-198. 

[2] Yan Z, Li P, Wang B, et al. Towards learning multi-domain 
crowd counting[J]. IEEE Trans. Circuits Syst. Video Technol, 
2021.Badrinarayanan, V., Kendall, A., Cipolla, R.: ‘Segnet: A 
deep convolutional encoder-decoder architecture for image 
segmentation’, IEEE Transactions on Pattern Analysis and 
Machine Intelligence, 2017, 39, pp. 2481–2495 

[3] Zou Z, Qu X, Zhou P, et al. Coarse to fine: Domain adaptive 
crowd counting via adversarial scoring 
network[C]//Proceedings of the 29th ACM International 
Conference on Multimedia. 2021: 2185-2194.Ghiasi,  

[4] Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image 
translation using cycle-consistent adversarial 
networks[C]//Proceedings of the IEEE international 
conference on computer vision. 2017: 2223-2232. 

[5] Wang Q, Gao J, Lin W, et al. Learning from synthetic data for 
crowd counting in the wild[C]//Proceedings of the IEEE/CVF 



 

37 

conference on computer vision and pattern recognition. 2019: 
8198-8207. 

[6] Gao J, Han T, Yuan Y, et al. Domain-adaptive crowd counting 
via high-quality image translation and density reconstruction[J]. 
IEEE transactions on neural networks and learning systems, 
2021. 

[7] Zhang Y, Zhou D, Chen S, et al. Single-image crowd counting 
via multi-column convolutional neural network [C] 
//Proceedings of the IEEE conference on computer vision and 
pattern recognition. 2016: 589-597. 

[8] Chan A B, Liang Z S J, Vasconcelos N. Privacy preserving 
crowd monitoring: Counting people without people models or 
tracking[C]//2008 IEEE conference on computer vision and 
pattern recognition. IEEE, 2008: 1-7ision, 2021, 129, (11), pp. 
3051–3068 

[9] Zhang Y, Zhou D, Chen S, et al. Single-image crowd counting 
via multi-column convolutional neural network [C] // 
Proceedings of the IEEE conference on computer vision and 
pattern recognition. 2016: 589-597. 

[10] Caesar, H., Uijlings, J., Ferrari, V.: ‘Coco-stuff: Thing and stuff 
classes in context’,Proceedings of the IEEE conference on 
computer vision and pattern recognition,2018, pp. 1209–1218