图像语义分割：从头开始训练DeepLab-LargeFOV

DeepLabv1安装及调试

前言：所谓DeepLab-LargeFOV就是DeepLab v1
论文：Semantic Images Segmentation With Deep Convolutional Nets And Fully Connected CRFs
介绍：在DCNN进行分割任务时，有两个瓶颈，一个是下采样所导致的信息丢失，通过带孔卷积的方法解决；另一个是空间不变性所导致的边缘不够准确，通过全连接的CRF解决。（CRF是可以通过底层特征进行分割的一个方法）。
核心工作：空洞卷积 + 如何降低计算量 + CRF作为后处理

1.调试环境

博主是在Ubuntu+docker条件下跑的，具体参数为：
GeForce GTX 1060 6GB
Ubuntu 16.04
Docker version 18.09.1
CUDA Version 8.0.61
NVIDIA Driver Version: 384.145

2.环境安装

1）docker镜像安装

$ docker pull rysenko/deeplab

挂载本地路径运行

$ sudo nvidia-docker run -it -v /home/ubuntu:/Deeplab rysenko/deeplab bash

2）PASCAL VOC2012 & Augmented PASCAL VOC2012下载

3）修改run.py中VOC2012数据集的路径

DATA_ROOT=' /Deeplab/datasets/VOCdevkit/VOC2012'

4）修改python/my_script/tools.py（原链接已失效）

#修改caffemodel的url链接
url = 'http://www.cs.jhu.edu/~alanlab/ccvl/init_models/vgg16_20M.caffemodel' 
#修改caffe路径
CAFFE_DIR='/Deeplab/DeepLab-Context/code'  #docker_caffe
CAFFE_BIN='.build_release/tools/caffe.bin'  #caffe.bin

5）matio安装：

$ pip install wget
$ wget http://sourceforge.net/projects/matio/files/matio/1.5.2/
$ cd matio-1.5.13
$ ./configure
$ make 
$ make check 
$ make install 
$ sudo ldconfig

6）Compile caffe
①修改Makefile.config
#需要把/opt/caffe中的文件拷贝到/DeepLab-Context/code后再编译

$ cp -r /opt/caffe/* /Deeplab/DeepLab-Context/code
$ cd /Deeplab/DeepLab-Context/code
$ cp Makefile.config.example Makefile.config  
$ vim Makefile.config  

#以下是对Makefile.config的修改

1. INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
2. LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnumake
3. # Uncomment the *_50 lines 

这里博主把Makefile.config部分贴出来供于参考（其余部分不做修改）

# Makefile.config
# ......
# CUDA architecture setting: going with all of them (up to CUDA 5.5 compatible).
# For the latest architecture, you need to install CUDA >= 6.0 and uncomment
# the *_50 lines below.
CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
		-gencode arch=compute_20,code=sm_21 \
		-gencode arch=compute_30,code=sm_30 \
		-gencode arch=compute_35,code=sm_35 \
		-gencode arch=compute_50,code=sm_50 \
		-gencode arch=compute_50,code=compute_50
# ......
# Whatever else you find you need goes here.
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnumake
# ......

②对hdf5进行链接

$ cd /usr/lib/x86_64-linux-gnu
$ ln -s libhdf5_serial.so.10.1.0 libhdf5.so
$ ln -s libhdf5_serial_hl.so.10.0.2 libhdf5_hl.so

否则出现以下错误

③编译caffe

$ make all -j16

3.Running the code

$ python run.py

#可修改/DeepLab-Context/voc12/config/vgg128_noup/train.prototxt中的batch_size
#以防出现out of memory 的情况，这里我们令batch_size = 4
#根据实际要求修改以下参数

4.运行效果

TRAIN

TEST

参考
https://blog.csdn.net/Xmo_jiao/article/details/77897109
https://zhuanlan.zhihu.com/p/34562137
https://blog.csdn.net/gzq0723/article/details/80153849