车辆检测--A Closer Look at Faster R-CNN for Vehicle Detection

A Closer Look at Faster R-CNN for Vehicle Detection

Intelligent Vehicles Symposium , 2016 :124-129

本文主要分析了 Faster R-CNN 对于车辆检测这个问题的性能表现，尝试了各种训练尺寸和测试图像尺寸

Examples from the KITTI car dataset



The network structure of Faster R-CNN



训练数据集和测试数据集



数据集上车辆尺寸分布图



B. What training scale is appropriate?

我们之间用 Faster R-CNN 在 KITTI 数据集上训练测试，训练输入图像尺寸较长的一边为 1000像素， only achieving 64.02% on the moderate car examples while state of the art results reported on the KITTI website are 90.03%

这个差距如何解释了？ 主要是降采样太多，车辆特征变小导致检测精度低

我们尝试了不同的训练图像尺寸



上图显示随着训练图像尺寸的增加，车辆检测精度是一直提升的。

However we used a training scale of 1500 for most of our analysis below for efficiency consideration.

C. Does the test scale matter?

测试图像的尺寸有没有影响了？



D. How many proposals are needed?



识别率上不去





11