OHEM安装运行(Training Region-based Object Detectors with Online Hard Example Mining)

OHEMgithub地址

https://github.com/abhi2610/ohem

结果可以重现的 论文结果 重现结果

Fast R-CNN (FRCN) VOC 07 trainval VOC 07 test 66.9 67.6

FRCN with OHEM VOC 07 trainval VOC 07 test 69.9 71.5

FRCN with OHEM VOC 07 trainval + 12 trainval VOC 07 test 74.6 75.5

最高结果，未复现 +M是多尺度，+B是bounding box regression

FRCN with OHEM, +M, +B VOC 07 trainval + 12 trainval VOC 07 test 78.9

一.软件要求

1.需要caffe和pycaffe支持

修改Makefile.config，WITH_PYTHON_LAYER := 1，USE_CUDNN := 1 可有可无

2.安装python包 cython, python-opencv, easydict, `yaaml’

3.代码中评价时用到了非官方的MATLAB版PASCAL VOC evaluation，代码中已经包含了。

二.正式安装

1.git clone –recursive https://github.com/abhi2610/ohem.git，之后成代码所在路径为OHEM_ROOT

2.编译Cython模块

cd OHEMROOT/libmake3.编译CaffeandpycaffecdOHEM_ROOT/caffe-fast-rcnn

make -j16 && make pycaffe

4.下载用OHEM算法在VGG16 and VGG_CNN_M_1024上训练的Fast R-CNN模型参数

cd $OHEM_ROOT

./data/scripts/fetch_fast_rcnn_ohem_models.sh

也可以直接从下面路径下载

http://graphics.cs.cmu.edu/projects/ohem/data/fast_rcnn_ohem_models.tgz

结果会存在$OHEM_ROOT/data/fast_rcnn_ohem_models文件夹中，形式类似fast_rcnn_ohem_models/TRAINING_SET/MODEL_FILE.

三.开始训练和测试模型

1.制作数据集

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

tar xvf VOCtrainval_06-Nov-2007.tar

tar xvf VOCtest_06-Nov-2007.tar

tar xvf VOCdevkit_08-Jun-2007.tar

VOCdevkit/ # development kitVOCdevkit/VOCcode/ # VOC utility code

VOCdevkit/VOC20072.创建PASCALVOC数据库的软连接cdOHEM_ROOT/data

ln -s VOCdevkitVOCdevkit2007四.下载ImageNet上预训练的模型cdOHEM_ROOT

./data/scripts/fetch_imagenet_models.sh

五.使用方法

用OHEM算法在voc_2007_trainval数据机上训练fast rcnn框架, 用下面的脚本 experiments/scripts/fast_rcnn_ohem.sh.

在experiments/scripts/目录下找耕读的脚本. 实验结果输出会写入在OHEMROOT/output，命令：cdOHEM_ROOT

./experiments/scripts/fast_rcnn_ohem.sh [GPU_ID] [NET] [–set …]

NET in {VGG16, VGG_CNN_M_1024} is the network arch to use

–set … allows you to specify fast_rcnn.config options, e.g.

–set EXP_DIR seed_rng1701 RNG_SEED 1701