Learning Spatiotemporal Features with 3D Convolutional Networks
设计初衷:带有全连接层的卷积网络适合迁移学习任务，但这种机遇图像特征提取的网络不适合时空相关性强的视频分析。
文章特点：提出一个容易理解、训练和使用的3D卷积网络。网络特点：降维处理压缩性好、适合时空相关性的视频分类中更有效。这个模型网络满足了一个有效的视频描述应该具有四个属性：开发一个泛化性好，适合各种类型的视频描述方法。
对于大规模的视频数据是高效的；可以高效计算的；简单有效的。
网络结构：3D卷积层和池化层.
应用
C3D能同时对外观和运动信息建模，具有高效、简单、紧凑的特点
C3D在Action recognition、Action Similarity Labeling、Scene and Object Recognition等任务上取得了较好的成绩。
代码简洁
import torch.nn as nn
from torch.nn import init
class c3d(nn.Module):
def __init__(self,num_class,init_weights=True):
super(c3d, self).__init__()
self.conv1a = nn.Conv3d(3, 64, kernel_size=3, padding=1)
self.conv2a = nn.Conv3d(64, 128, kernel_size=3, padding=1)
self.conv3a = nn.Conv3d(128, 256, kernel_size=3, padding=1)
self.conv3b = nn.Conv3d(256, 256, kernel_size=3, padding=1)
self.conv4a = nn.Conv3d(256, 512, kernel_size=3, padding=1)
self.conv4b = nn.Conv3d(512, 512, kernel_size=3, padding=1)
self.conv5a = nn.Conv3d(512, 512, kernel_size=3, padding=1)
self.conv5b = nn.Conv3d(512, 512, kernel_size=3, padding=1)
self.pool1 = nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2))
self.pool2 = nn.MaxPool3d(kernel_size=2, stride=2)
self.pool3 = nn.MaxPool3d(kernel_size=2, stride=2)
self.pool4 = nn.MaxPool3d(kernel_size=2, stride=2)
self.pool5 = nn.MaxPool3d(kernel_size=2, stride=2)#, padding=(0, 1, 1)
self.fc6 = nn.Linear(4608, 4096)
self.fc7 = nn.Linear(4096, 4096)
self.out = nn.Linear(4096, num_class)
self.relu = nn.ReLU()
self.softmax = nn.Softmax()
if init_weights:
self._initialize_weights()
def forward(self, x):
x = self.conv1a(x)
x = self.relu(x)
x = self.pool1(x)
x = self.conv2a(x)
x = self.relu(x)
x = self.pool2(x)
x = self.conv3a(x)
x = self.relu(x)
x = self.conv3b(x)
x = self.relu(x)
x = self.pool3(x)
x = self.conv4a(x)
x = self.relu(x)
x = self.conv4b(x)
x = self.relu(x)
x = self.pool4(x)
x = self.conv5a(x)
x = self.relu(x)
x = self.conv5b(x)
x = self.relu(x)
x = self.pool5(x)
x = x.view(x.size(0), -1)
x = self.fc6(x)
x = self.relu(x)
x = self.fc7(x)
x = self.relu(x)
res = self.out(x)
# if you use CrossEntropyLoss, you don't need to add softmax in network
# res = self.softmax(x)
return res
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv3d):
init.xavier_uniform_(m.weight)
if m.bias is not None:
init.constant_(m.bias,0)
elif isinstance(m, nn.Linear):
init.xavier_uniform_(m.weight)
init.constant_(m.bias, 0)
其他资源：
https://blog.csdn.net/xiaofei0801/article/details/72896354
https://blog.csdn.net/zzmshuai/article/details/84866514
https://github.com/MRzzm/action-recognition-models-pytorch/blob/master/3DCNN/c3d/c3d.py
