.png){kind=icon}
返回
SCU-Counting: A large-scale benchmark dataset for multi-class object counting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.trc.2024.104608.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Object counting has been widely studied because of its practicality. However, most currently published object counting datasets contain only a single class of objects, such as crowd or vehicle, which is challenging to meet the usage needs of tasks, such as traffic control. For practical applications, a multi -class object counting (MCOC) dataset that includes transports is increasingly needed. Therefore, this study collected a large-scale MCOC dataset from the Web and live -viewing, called the Soochow university counting (SCU-Counting) dataset. Images from the Web are licensed for free modification, sharing and using for all creative commons. SCU-Counting consists of 2521 images, where 108 images are empty scenes and others are annotated, in a total of 274,199 annotated instances (195,018 humans, 38,831 cars, 2503 trucks, 6445 buses, 10,964 tricycles, 4600 e -bikes, 9729 e -scooters, 3900 bicycles, 2169 motorcycles). Except for empty scenes, each image contains at least one type of objects. In addition, this study proposes a convolutional neural network -based multi -aware feature aggregation network (MAFA-Net) for MCOC. MAFA-Net first uses a front-end module to extract shallow image features, and then designs a multi -aware module that can aggregate contextual, channel and spatial information of features extracted by the front-end. Additionally, a class adaptive weight assignment (CAWA) loss function is developed based on the focal loss function, which can regulate the sample balance between classes in MCOC tasks. We conducted comparative experiments on five public counting datasets to evaluate the effectiveness of our proposed MAFA-Net. In addition, we conducted extensive experiments on the SCU-Counting dataset to validate the counting and target localization abilities of MAFA-Net. For the counting task, MAFA-Net achieves the best MAE (mean absolute error) of 20.08 and MSE (mean squared error) of 46.11, which is 1.84 and 1.75 lower in MAE and MSE than the second best method, respectively. For the localization task, MAFA-Net reaches 90.3% normalized average precision (nAP) at the level of 0.5 and 67.5% average nAP at the level set of {0 . 05 , 0 . 1 , ... , 0 . 50}, which suggests that the localization performance of MAFA-Net can meet the requirement of applications. In summary, experimental results demonstrate the usability of the SCU-Counting dataset and the superiority of MAFA-Net.
Keyword:
Crowd counting
Multi-class object counting
Object counting dataset
Multi-aware feature aggregation
期刊
IF:
7.9
论文数:
4.7K
被引数:
3.2W
机构
引用论文
Exploiting Vibrational Strong Coupling to Make an Optical Parametric Oscillator Out of a Raman Laser
OAAI
OAAI分享收藏T2CNN: a novel method for crowd counting via two-task convolutional neural networkT2CNN: 一种基于双任务卷积神经网络的人群计数新方法.png)
VISUAL COMPUTER
VISUAL COMPUTERIF2.9
PREAI
分享收藏Electrochemically assisted micro localized grafting of aptamers in a microchannel engraved in fluorinated thermoplastic polymer Dyneon THVRSC Advances
RSC AdvancesIF0
PREAI
分享收藏
查看更多内容