提取C3D视频特征(官方文档&实践)

C3D Introduction

卷积神经网络（CNN）近年被广泛应用于计算机视觉中，包括分类、检测、分割等任务。这些任务一般都是针对图像进行的，使用的是二维卷积（即卷积核的维度为二维）。而基于视频的问题，提特征的方法主要分为双流和C3D两个分支，目前C3D衍生出P3D，I3D等等，这里只介绍最早的C3D caffe版提取特征的步骤以及遇到的问题。

C3D 用caffe实现
官网
github

C3D Installation

installation

C3D User Guide

官方User Guide

Du Tran (Last modified Mar 20, 2017)
C3D-v1.1 is released with new models (Mar 01, 2017).
No documentation for v1.1 yet, but some examples for feature extraction, training, and fine-tuning are provided. 

文档介绍了C3D 1.0的用法，至于1.1，只有examples。

I.C3D Feature Extration

1. 安装好C3D，下载预训练的模型，保存到

   YOUR_PATH_TO_C3D/C3D-master/C3D-v1.0/examples/c3d_feature_extraction

2. change dir to

   YOUR_PATH_TO_C3D/C3D-master/C3D-v1.0/examples/c3d_feature_extraction

3. Run:

   sh c3d_sport1m_feature_extraction_frm.sh
   or
   sh c3d_sport1m_feature_extraction_video.sh