论文解读《Deep People Counting in Extremely Dense Crowds》ACM MM2015

Deep People Counting in Extremely Dense Crowds

极度密集的深度人群计数

Chuan Wang, Hua Zhang, Liang Yang, Si Liu, Xiaochun Cao

 

摘要:

1. 问题:

缺少训练样本、严重遮挡、杂乱的背景、视角变化

2. 提出端到端的深度CNN回归模型对极其密集人群计数。

3. 特性:

基于CNN的深度模型,可以自动学习有效的特征计数

to weaken influence of background like buildings and trees, we purposely enrich the training data with expanded                              negative samples whose ground truth counting is set as zero.

为了减弱建筑物、树木等背景的影响,我们特意使用扩大的负样本来丰富训练数据,并将其ground truth count(真实                      人数)设为0。

方法:

1. Data Collection and Preparation

51幅图,平均每张731个人,95-3714。最后数据增强,裁剪、翻转生成6414个patches输入网络。

论文解读《Deep People Counting in Extremely Dense Crowds》ACM MM2015

2. Convolutional Neural Network for Regression

 

论文解读《Deep People Counting in Extremely Dense Crowds》ACM MM2015

 

论文解读《Deep People Counting in Extremely Dense Crowds》ACM MM2015

pj、gj是预测的人数和真实的人数

3. Negative Samples

709个负样本,学习树木、建筑等作为人群图像的背景特征。这些图片的真实人数为0。

 

实验:

We use standard evaluation criteria, mean and deviation of Absolute Difference (AD) and these of Normalized Absolute Difference (NAD), to quantify the performance.

使用AD和NAD作为性能指标,在UCF_CC_50上实验。

论文解读《Deep People Counting in Extremely Dense Crowds》ACM MM2015