人群分析--Beyond Counting: Comparisons of Density Maps for Crowd Analysis Tasks

Beyond Counting: Comparisons of Density Maps for Crowd Analysis Tasks - Counting, Detection, and Tracking
https://arxiv.org/abs/1705.10118

基于人群密度图估计的人群分析，这里我们主要对比各种 CNN-based density estimation methods 用于 counting, detection, and tracking，我们得到的结论是
high-quality density maps 对于 detection 和 tracking 都是有帮助的。

基于密度图估计的计数思想最早是由文献【3】提出的 Maximum Excess over Sub Array (MESA) distance。 随后各种 CNN 来估计密度图

本文主要考虑两个CNN网络生成 full-resolution density maps：CNN-pixel 和 FCNN-skip

对于 CNN-pixel，我们使用一个 classic CNN regressor for pixel-wise density prediction 也就是说用滑动窗口的方式来得到每个像素的预测结果
given an image patch, predict the density at the center pixel. The full-resolution density map is obtained using a sliding window to obtain density values for all pixels inside the ROI.

对于 Fully Convolutional Architecture： hole convolution 的效果要比 FCNN-skip 差

C. Detection from Density Maps
根据密度图来做检测，首先是文献【11】提出的

D. Improving Tracking with Density Maps
用密度图来改善跟踪

Experiments
实验使用的数据库：

UCSD dataset

Summary
In summary, CNN-pixel performs best on various metrics, including compactness, localization MAE, and temporal smoothness. This suggests that using full-resolution perpixel prediction of the density map can improve detection and tracking. In contrast, reduced-resolution maps that require either fixed upsampling (CNN-patch and MCNN) or learned upsampling-convolution (FCNN-skip and MCNN-up) had worse quality metrics, resulting in poorer detection/tracking accuracy. The downsampling operations in the CNN obfuscate the true position of the people in the reduced-resolution maps. This loss of spatial information can only be partially compensated using learned upsampling and skip connections.

However, dense prediction suffers from higher computational complexity, compared to fully-convolutional networks

一句话： CNN-pixel 效果好但是计算量大