VOC-release4和VOC-release5检测速度测试

PC：Inter core(TM)i5-2400 CPU3.10GHz RAM4.00GB

检测图片大小：640*480

加载model：INRIA样本库训练出的单组件8个子部件的model

每次的检测时间会有微小浮动，与电脑CPU状态和图片内容有关：检测时移动鼠标都会对检测时间有很大影响；目标（人）数目不同时，保存检测窗位置画出最终检测窗的运算量会有差别。由于浮动不大，所以我直接取了其中一次，没有做多次平均。

release4：

0)multithreaded convolution using SSE        fconvsse.cc          Time:1.672259 seconds

1)multithreaded convolution using blas         fconvblasMT.cc   Time:4.588164 seconds

2)multithreaded convolution without blas      fconvMT.cc          Time:2.407935 seconds

3)convolution using blas                                fconvblas.cc        Time:7.052188 seconds

4)blas convolution, very compatible              fconv.cc               Time:4.588164 seconds

 

release5:

0）multithreaded convolution using SSE           and  0）multithreaded convolution 

fconvsse.cc and conv_var_dim_MT.cc                      Time：3.782209 seconds

1）multithreaded convolution                            and 0）multithreaded convolution

fconv_var_dim_MT.cc and conv_var_dim_MT.cc     Time：6.277308 seconds

2）basic convolution, very compatible               and  0）multithreaded convolution 

conv_var_dim.cc and conv_var_dim_MT.cc              Time：15.491376 seconds

 

0）multithreaded convolution using SSE           and  1）single-threaded convolution

fconvsse.cc and conv_var_dim.cc                              Time：3.598026 seconds

1）multithreaded convolution                             and  1）multithreaded convolution 

fconv_var_dim_MT.cc and conv_var_dim.cc              Time：6.285950 seconds

2）basic convolution, very compatible               and  1）multithreaded convolution 

fconv_var_dim.cc and conv_var_dim.cc                     Time：15.473099 seconds

5比4的最大改进在训练代码，迭代收敛更快训练时间更短：Training is done in memory；Optimization improvements (faster convergence)； Weak-label structural SVM (wl-ssvm)。