Deep Learning Strong Parts for Pedestrian Detection

ICCV 2015

本文主要利用深度学习结合 part model 得到一个 DeepParts 来解决 行人检测 中的 遮挡问题。

DeepParts has four main contributions：
1）通过数据驱动，自动选择若干互补的局部模型
2) we are the first to extensively explore how single part detector and
their ensemble based on ConvNets contribute to pedestrian detection
3) We propose a novel method to handling proposal shifting problem
4) We show that with complementary part selection, a new state-of-the-art miss
rate of 11.89% can be achieved on the Caltech reasonable set

先看看下图：

part model 怎么选择？ the part selection is determined by data and the effectiveness of the part pool can be fully explored.

2 Training Part Detectors
我们首先构建一个 part pool, 然后 对每个 part 训练一个检测器，针对proposal windows 偏移问题提出一个解决方法，最后综合所有检测器的分数，得到整个行人检测结果。

2.1. Part Pool
我们一共选了 45个 prototypes。Two examples regarding the parts of head-left-shoulder and leg are shown

2.2. Training

这里我们尝试了三个模型，三个预训练策略
Three deep models are AlexNet [15], Clarifai [39], and GoogLeNet [30]
Three pre-training strategies include：
(1) 参数高斯分布随机初始化，无 预训练 (strategy 1)，
(2) ImageNet 预训练 (strategy 2)，
(3) 输入 ImageNet 图像中的 object 块预训练 (strategy 3)

2.3. Handle Shifting in Deep Model

4 Experiments