Multiscale Aggregate Networks with Dense Connections for Crowd Counting

Abstract

The most advanced method for crowd counting uses a fully convolutional network that extracts image features and then generates a crowd density map. However, this process often encounters multiscale and contextual loss problems. To address these problems, we propose a multiscale aggregation network (MANet) that includes a feature extraction encoder (FEE) and a density map decoder (DMD). The FEE uses a cascaded scale pyramid network to extract multiscale features and obtains contextual features through dense connections. The DMD uses deconvolution and fusion operations to generate features containing detailed information. These features can be further converted into high-quality density maps to accurately calculate the number of people in a crowd. An empirical comparison using four mainstream datasets (ShanghaiTech, WorldExpo’10, UCF_CC_50, and SmartCity) shows that the proposed method is more effective in terms of the mean absolute error and mean squared error. The source code is available at https://github.com/lpfworld/MANet.

Neural Network-Based Crowd Counting Systems: State of the Art, Challenges, and Perspectives

“…Attentive Feature Refinement [57] block in the encoder to adaptively extract multi-scale features. The decoder's Non-local Fusion block aggregates multi-scale information from several layers at lower computation costs using selfattention mechanism.…”

“…Another metric mean error is proposed in [57] to measure the variance of ground truths and it is defined as follows:…”

SENetCount: An Optimized Encoder-Decoder Architecture with Squeeze-and-Excitation for Crowd Counting

“…The split-transform-merge technique is anticipated to access the representational capacity of high-density and largescaled layers but at an appreciably lower calculating cost. [10] 2018 ECCV Multicolumn Whole image Multitask SAAN [11] 2019 WACV Multicolumn Whole image Single-task SPN [12] 2019 WACV Single-column Whole image Single-task LA-Batch [13] 2022 TPAMI Single-column Patch Multi-task MANet [14] 2021 CIAN Single-column Whole image Single-task LSC-CNN [15] 2021 TPAMI Multicolumn Whole image Multitask (include detection) AutoScale [16] 2022 IJCV Multicolumn Patch Multitask (include detection)…”

“…Referring to existing works, we evaluate our method with both the mean absolute error (MAE) and the root mean squared error (RMSE), which are adopted to assess the capability of the proposed methods [9][10][11][12][13][14]. The MAE demonstrates the accuracy while the RMSE reflects the robustness, and a lower value means better capability.…”

“…iments are divided into multicolumn-based methods MCNN, ic-CNN, and SAAN[9][10][11], single-column-based methods SPN, LA-Batch, and MANet[12][13][14], and detection-based methods AutoScale and LSC-CNN[15,16].…”