PoseNet文章复现记录

╰+哭是因爲堅強的太久メ 2022-06-01 12:39 205阅读 0赞

## 下载链接 ##

1.  文章：[PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization][PoseNet_ A Convolutional Network for Real-Time 6-DOF Camera Relocalization]
2.  代码：[alexgkendall/caffe-posenet][alexgkendall_caffe-posenet]
3.  数据集：[地址][Link 1]
4.  模型参数：[地址][Link 2]

## 编译caffe-posenet ##

1.  参考上一篇文章：[Ubuntu下Caffe安装记录][Ubuntu_Caffe]
2.  问题解决：
    
        In file included from ./include/caffe/util/device_alternate.hpp:40:0,
                            from ./include/caffe/common.hpp:19,
                             from src/caffe/syncedmem.cpp:1:
            ./include/caffe/util/cudnn.hpp: In function ‘void caffe::cudnn::createPoolingDesc(cudnnPoolingStruct**, caffe::PoolingParameter_PoolMethod, cudnnPoolingMode_t*, int, int, int, int, int, int)’:
            ./include/caffe/util/cudnn.hpp:127:41: error: too few arguments to function ‘cudnnStatus_t cudnnSetPooling2dDescriptor(cudnnPoolingDescriptor_t, cudnnPoolingMode_t, cudnnNanPropagation_t, int, int, int, int, int, int)’
                     pad_h, pad_w, stride_h, stride_w));
                                                     ^
            ./include/caffe/util/cudnn.hpp:15:28: note: in definition of macro ‘CUDNN_CHECK’
         cudnnStatus_t status = condition; \
    
    方法：
    
     *  将./include/caffe/util/cudnn.hpp 换成最新版的caffe里的cudnn的实现，即相应的cudnn.hpp.
     *  将./include/caffe/layers里的，所有以cudnn开头的文件，例如cudnn\_conv\_layer.hpp。 都替换成最新版的caffe里的相应的同名文件。
     *  将./src/caffe/layer里的，所有以cudnn开头的文件，例如cudnn\_lrn\_layer.cu， cudnn\_pooling\_layer.cpp， cudnn\_sigmoid\_layer.cu 都替换成最新版的caffe里的相应的同名文件。

## 切换Pycaffe环境 ##

# 因为我有俩个caffe，分别是官方的和posenet的，分别都编译了，
    # import caffe时需指定导入哪一个。
    
    sudo gedit ~/.bashrc
    # 更改完caffe路径或者caffe-posenet路径
    source ~/.bashrc

## 数据集介绍 ##

1.  以KingsCollege举例，下面包含8个序列的场景图片。2，3，7是作为测试集（dataset\_test.txt），1，4，5，6，8作为训练集（dataset\_train.txt）。
    
    ![这里写图片描述][SouthEast]
2.  下载的模型文件如下图。
    
    ![这里写图片描述][SouthEast 1]

## 创建lmdb数据集 ##

1.  修改caffe-posenet/posenet/scripts下的create\_posenet\_lmdb\_dataset.py文件。一共三处，分别是1，11，12行；
    
    ![这里写图片描述][SouthEast 2]
2.  安装lmdb；
    
        sudo pip install lmdb
3.  创建数据集:python posenet/scripts/create\_posenet\_lmdb\_dataset.py。
    
    ![这里写图片描述][SouthEast 3]
4.  创建完的数据集生成在了caffe-posenet目录下，下图的posenet\_dataset\_lmdb文件夹；
    
    ![这里写图片描述][SouthEast 4]

## 创建均值文件 ##

1.  新建一个名为“create\_posenet\_mean.sh”的文件；
    
        #!/usr/bin/env sh 
        set -e
        PATH=./data
        DATA=./data
        DBTYPE=lmdb
        echo "Computing image mean..."
        ./build/tools/compute_image_mean -backend=$DBTYPE \
          $PATH/posenet_dataset_$DBTYPE $PATH/mean.binaryproto
        echo "Done."
    
    ![这里写图片描述][SouthEast 5]
2.  将上一步生成的lmdb数据集放到caffe-posenet/data目录下；
3.  运行create\_posenet\_mean.sh，得到mean.binaryproto文件；
    
        # 添加可执行权限
        
        chmod 777 posenet/scripts/create_posenet_mean.sh
        ./posenet/scripts/create_posenet_mean.sh
    
    ![这里写图片描述][SouthEast 6]
4.  将数据和均值文件改名，前缀加train\_；  
    ![这里写图片描述][SouthEast 7]

## 修改网络配置 ##

1.  修改“train\_kingscollege.prototxt”中name为data，phase为TEST和TRAIN的层的source和mean\_file的路径：
    
    ![这里写图片描述][SouthEast 8]  
    ![这里写图片描述][SouthEast 9]
2.  安装python-tk：sudo apt-get install python-tk。

## 直接测试 ##

1.  输入命令：
    
        sudo python posenet/scripts/test_posenet.py --model models/PoseNet_rawmodel/train_kingscollege.prototxt --weights models/PoseNet_rawmodel/weights_kingscollege.caffemodel --iter 8
2.  测试结果：
    
    ![这里写图片描述][SouthEast 10]

## 结果 ##

1.  误差没有文章效果好，是为什么呢？

## 参考文章 ##

1.  [PoseNet运行详细过程][PoseNet]
2.  [PoseNet简介及运行][PoseNet 1]

[PoseNet_ A Convolutional Network for Real-Time 6-DOF Camera Relocalization]: https://arxiv.org/abs/1505.07427
[alexgkendall_caffe-posenet]: https://github.com/alexgkendall/caffe-posenet
[Link 1]: http://mi.eng.cam.ac.uk/projects/relocalisation/#dataset
[Link 2]: http://mi.eng.cam.ac.uk/~agk34/resources/PoseNet.zip
[Ubuntu_Caffe]: http://blog.csdn.net/bskfnvjtlyzmv867/article/details/79099289
[SouthEast]: /images/20220601/f4794e50ebe3472abc9627f85e8883cc.png
[SouthEast 1]: /images/20220601/192ad2fca8bd4652b73c36cec7d577e4.png
[SouthEast 2]: /images/20220601/1a764aa82b3945898ea3f25a34d01b18.png
[SouthEast 3]: /images/20220601/00c2d1f112754f578d30e4d03252ebb0.png
[SouthEast 4]: /images/20220601/3da3efd8ec4c42c0b6be6f4c4d1c73e7.png
[SouthEast 5]: /images/20220601/44b8746650754c638ca82de0d3c4939b.png
[SouthEast 6]: /images/20220601/681a3cbfbe804dcbb49d469d5fc169b9.png
[SouthEast 7]: /images/20220601/4b40efb4364845098c8fdbd82bcf25ba.png
[SouthEast 8]: /images/20220601/3c17dc8172074ecebe12ddf6257ae522.png
[SouthEast 9]: /images/20220601/d1cef5f99c8b475db2553f3f0ff66692.png
[SouthEast 10]: /images/20220601/63893a83e7624a3fb79fad3cd66480f7.png
[PoseNet]: http://blog.csdn.net/xuzhongxiong/article/details/54565105
[PoseNet 1]: http://blog.csdn.net/u013832707/article/details/53704093