Learning Complexity-Aware Cascades for Deep Pedestrian Detection

ICCV 2015

本文主要介绍了一种考虑计算复杂度的级联器用于检测行人，将CNN特征嵌入级联器中。 对于不同的特征，计算量是不一样的。在级联器中，我们希望使用简单的特征来排除大部分可能位置，对于少数通过简单特征判断的位置，我们使用复杂的特征来确认。

3 Complexity-Aware Cascade Training
3.1. AdaBoost
AdaBoost 由若干若分类器组成一个强分类器。

3.2. Complexity-Aware Learning
将误差函数定义为两部分，classification accuracy and complexity

3.3. Embedded Cascade

3.4. Cascade Boosting

The optimal step size for the update is

The cascade predictor is finally updated with

This boosting algorithmis denoted the complexity aware cascade training
(CompACT) boosting algorithm

4 Pedestrian Detection
4.1. Feature Pools of Variable Complexity
对于不同的特征，计算量是不一样的。在级联器中，我们希望使用简单的特征来排除大部分可能位置，对于少数通过简单特征判断的位置，我们使用复杂的特征来确认。

Pre-computationis useful for low-complexity features, complex features should be evaluated as necessary. We refer to the formeras pre-computed features and the latter as computed just-in-time (JIT).

4.1.1 Pre-computed Features
Our pre-computed feature set consists of ACF [4], mostly due to its computational efficiency. Following [4], we extract 10 LUV+HOG channels.

4.1.2 Just-in-time Features
SS: The self-similarity (SS) features of [28]
CB: Checkerboard features (CB) are the result of convolving the ACF channels with a set of checker board filters. [37]
LDA: Locally decorrelated HOG features
CNN： is a smaller version of the popular model of [17], with five convolutional layers and one fully connected layer