Comparing the differences between Faster RCNN and RPN+BF in pedestrain detection
By zyq&cyq
Introduction
行人检测具有极其广泛的应用:智能辅助驾驶,智能监控,行人分析以及智能机器人等领域。随着深度学习的性能的优越性,将深度学习的方法应用到行人中以提高检测准确率。本工程分别采用Faster R-CNN和RPN+BF网络,对Caltech数据集进行训练和测试,并比较两者的结果。
This code has been tested on Ubuntu 16.04 with MATLAB 2014b and CUDA 7.5.
Citing
RPN+BF
@article{zhang2016faster,
title={Is Faster R-CNN Doing Well for Pedestrian Detection?},
author={Zhang, Liliang and Lin, Liang and Liang, Xiaodan and He, Kaiming},
journal={arXiv preprint arXiv:1607.07032},
year={2016}
}
Faster R-CNN
@article{ren15fasterrcnn,
Author = {Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun},
Title = {{Faster R-CNN}: Towards Real-Time Object Detection with Region Proposal Networks},
Journal = {arXiv preprint arXiv:1506.01497},
Year = {2015}
}
Requirements
-
Caffe
build for RPN+BF (see here)- If the mex in 'external/caffe/matlab/caffe_faster_rcnn' could not run under your system, please follow the instructions on our Caffe branch to compile and replace the mex.
-
MATLAB
-
GPU: Titan X, K40c, etc.
How to build and run
-
Download the special caffe vision for this project(see here), and follow the readme.md in it to build and run.
-
Download the annotations and videos in Caltech Pedestrian Dataset and put them in the three folder (videos|res|annotations) under ./RPN_BF/external/code3.2.1/data-USA and ./faster_rcnn_caltech/external/code3.2.1/data-USA.
-
The ./faster_rcnn_caltech include the code of faster rcnn on caltech datasets, follow the readme.md to make sure it perform well. Start MATLAB from the repo folder, and Run
script_faster_rcnn_caltech.m
to train and test the faster rcnn on Caltech,script_fast_rcnn_caltech_eval.m
to evaluate the result after train and test. -
The ./RPN_BF include the code of RPN+BF on caltech datasets, follow the readme.md to make sure it perform well, Start MATLAB from the repo folder, and Run
script_rpn_pedestrian_VGG16_caltech
to train and test the RPN model on Caltech, Runscript_rpn_bf_pedestrian_VGG16_caltech
to train and test the BF model on Caltech (the evaluation result is included in the test). -
Hopefully it would give the evaluation results.
Experiment results
Faster RCNN
In addition, we have raised the mr to 30% for Faster RCNN on the caltech datasets.