How Far are We from Solving Pedestrian Detection?

原文链接 http://blog.****.net/cv_family_z/article/details/52119644


CVPR 2016 我们离解决行人检测问题到底还有多远?

How Far are We from Solving Pedestrian Detection? 
项目网页:https://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/people-detection-pose-estimation-and-tracking/how-far-are-we-from-solving-pedestrian-detection/

Matlab code: https://bitbucket.org/shanshanzhang/code_filteredchannelfeatures

行人检测问题一直是一个比较热门的研究话题,行人检测最近几年进步比较大,那么还有多大的提升空间了?最近建立了一个 human baseline(人自己去看,来框出行人),发现大约目前最先进的算法与人工检测差 10%。

How Far are We from Solving Pedestrian Detection?

于是我们分析了一下目前算法主要存在的问题,找出改进的空间:1) high confidence false positives 一个主要因素是 localisation ,这里我们通过改进 training set alignment quality 来解决;2) 背景的干扰,通过深入分析CNN网络来改进

这里我们首先建立了一个 human baseline, 这是我们的终极目标,达到或者超过人工检测率。

How Far are We from Solving Pedestrian Detection?

How Far are We from Solving Pedestrian Detection?

3.2. Failure analysis

紧接着我们分析了一下 当前最好的算法检测失败的原因 
How Far are We from Solving Pedestrian Detection? 
How Far are We from Solving Pedestrian Detection?

Conclusion: For most top performing methods localisation and background-vs-foreground errors have equal impact on the detection quality. They are equally important.

3.3. Improved Caltech-USA annotations 
原来的数据库真值有不太准确的地方,我们对此进行了改进。

4.1. Impact of training annotations 
How Far are We from Solving Pedestrian Detection?

4.2. Convnets for pedestrian detection

How Far are We from Solving Pedestrian Detection?

How Far are We from Solving Pedestrian Detection?

Conclusion :CNN网络在图像分类和广义目标检测问题显示出很强的能力,但是针对小目标检测定位问题表现的不是很好,加入了Bounding box regression (and 
NMS)有所改善,但是 背景的干扰仍然是检测失败主要的原因。