Abstract | ||
---|---|---|
Crowd counting is a challenging task due to occlusions, continuous scale variation of target and perspective distortion. The existing density-based approaches usually utilize deep convolutional neural network (CNN) to regress a density map from deep level features and obtained the counts. However, the best results may be obtained from the features of lower level instead of deep level. It is mainly due to the overfitting that degrades the adaptability towards the continuous scale variation of target. To address the issue of overfitting, a novel approach, called gated cascade multi-stage regression network (GC-MRNet), was proposed. It aims to maintain the adaptability towards scale variation of target and generate higher accuracy estimated density maps. Firstly, the dense scale network (DSNet) was used as the backbone and multi-stage regression was employed to achieve different density map regressors in different levels. Then, the features derived from the density map were cascaded to assist generating a higher quality density map in next stage. Finally, the gated blocks were designed to achieve the controllable information interaction between cascade and backbone. Extensive experiments were conducted on the ShanghaiTech, UCF-QNRF and UCF-CC-50 datasets. The results demonstrated significant improvements of GC-MRNet, almost over the state-of-the-art on ShanghaiTech Part A. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1007/978-3-030-86340-1_5 | ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2021, PT II |
Keywords | DocType | Volume |
Crowd counting, Kernel density estimation, Feature extraction, Cascade stages, Gated block | Conference | 12892 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 6 |
Name | Order | Citations | PageRank |
---|---|---|---|
Ying Shi | 1 | 0 | 0.34 |
Jun Sang | 2 | 40 | 12.62 |
Jinghan Tan | 3 | 0 | 0.68 |
Zhongyuan Wu | 4 | 1 | 2.05 |
Bin Cai | 5 | 3 | 1.43 |
Nong Sang | 6 | 475 | 72.22 |