CAFN: The Combination of Atrous and Fractionally Strided Convolutional Neural Networks for Understanding the Densely Crowded Scenes
The task to estimate crowd count in highly clustered scenes is extremely challenged on account of variable scales with non-uniformity. This paper aims to develop a simple but valid method that concentrates on predicting the density map accurately. We proposed a combination of atrous and fractionally strided convolutional neural network (CAFN), which is merely constituted by two components: an atrous convolutional neural network as the frontend for 2D features extraction which utilizes dilated kernels to deliver larger receptive fields and to lessen the network parameters, a fractionally strided convolutional neural network for the back-end to lower the loss of details during down-sampling. CAFN is an easy-trained model because of its unadulterated convolutional structure. We demonstrated CAFN on three datasets (Shanghai Tech dataset A and B, UCF_CC_50) and deliver satisfactory performance. Additionally, CAFN achieves lower Mean Absolute Error (MAE) on Shanghai Tech A (MAE = 100.8), UCF_CC_50 (MAE = 305.3) while the experiment results reveal that the proposed model can effectively lower estimation errors when compared with previous methods.
CAFN Crowd density estimation Atrous convolutions Fractionally strided convolutions
Lvyuan Fan Minglei Tong Min Li
Shanghai University of Electric Power,Shanghai 200082,Peoples Republic of China
国际会议
广州
英文
297-307
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）