看到一篇不需要预训练，就可以直接进行联合训练直接检测的文章。
代码:
github caffe代码
github tensorflow代码
caffe结构图:点这里
我下载的是tensorflow版本的。
论文简介：
这篇文章是按照SSD方式进行目标检测的。结构的对比是按照SSD来办的。这里的Plain Connection 就是SSD，而Dense Connection就是DSOD了。
-
图中左边的就是SSD，融合了多个featu map层的特征进行预测，图中的粗的黑线就是把这张特征图传到最后，虚线是说明在这方框中进行的操作，就是conv（1x1+3x3） + （pooling strde= 2）。在模型尾端进行目标检测。
-
图中右边即使DSOD，与SSD一样融合了多个featu map层的特征进行预测，这里其实可以看到是两个网络并行。蓝色的特征图是一个网络，右边的框是一个网络这个网络只有6层。蓝色就是基于DenseNets的网络，而右边的框的输入不仅来自上层网络，而且会加上左边蓝色特征图网络的输出，这里使用contat进行相加。
-
检测： 按照SSD的思路，就在不同尺度的使用不同尺度的box，在feature map上进行box的预测和classes分类。
好了这里的模型介绍结束
下载了这里的tensorflow版的代码，发现这里和论文不一样。
def inference(inputs, is_train, reuse):
W_init = tf.contrib.layers.xavier_initializer()
with tf.variable_scope("model", reuse=reuse):
tl.layers.set_name_reuse(reuse)
net = InputLayer(inputs, name='input')
net = Conv2d(net, 64, (3, 3), (2, 2), padding='SAME',
W_init=W_init, name='stem1')
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='stem1_bn')
net = Conv2d(net, 64, (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='stem2')
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='stem2_bn')
net = Conv2d(net, 128, (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='stem3')
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='stem3_bn')
net = MaxPool2d(net, filter_size=(2, 2), strides=(2, 2), name='stem3_pool')
net = denseblock(net, blocknum=6, step=48, firstchannel=192, is_train=is_train, name='denseblock0', reuse=reuse)
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='denseblock0_bn')
net = Conv2d(net, 416, (1, 1), (1, 1), padding='SAME',
W_init=W_init, name='denseblock0_cnn')
net = MaxPool2d(net, filter_size=(2, 2), strides=(2, 2), name='denseblock0_pool')
net = denseblock(net, blocknum=8, step=48, firstchannel=192, is_train=is_train, name='denseblock1', reuse=reuse)
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='denseblock1_bn')
net = Conv2d(net, 800, (1, 1), (1, 1), padding='SAME',
W_init=W_init, name='denseblock1_cnn')
# 第一次的特征输出，没有相加其他的值
netfirst=BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='feature_first_bn')
net = MaxPool2d(net, filter_size=(2, 2), strides=(2, 2), name='denseblock2_pool1')
net = denseblock(net, blocknum=8, step=48, firstchannel=192, is_train=is_train, name='denseblock2', reuse=reuse)
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='denseblock2_bn')
net = Conv2d(net, 1184, (1, 1), (1, 1), padding='SAME',
W_init=W_init, name='denseblock2_cnn')
net = denseblock(net, blocknum=8, step=48, firstchannel=192, is_train=is_train, name='denseblock3', reuse=reuse)
net = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='denseblock3_bn')
net = Conv2d(net, 256, (1, 1), (1, 1), padding='SAME',
W_init=W_init, name='denseblock2_cnna')
netpl=MaxPool2d(netfirst, filter_size=(2, 2), strides=(2, 2), name='First_pool')
netpl=BatchNormLayer(netpl, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='First_bn')
netpl = Conv2d(netpl, 256, (1, 1), (1, 1), padding='SAME',
W_init=W_init, name='denseblock2_cnnb')
net=ConcatLayer([net,netpl],-1,"Second_Cat")
# 第二次的特征融合
netsecond = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu, name='feature_second_bn')
net = denseblockpl(net, step=256, firstchannel=256, is_train=is_train, name='denseplz1', reuse=reuse)
# 第三次的特征融合
netthird = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu,
name='feature_third_bn')
net = denseblockpl(net, step=128, firstchannel=128, is_train=is_train, name='denseplz2', reuse=reuse)
# 第四次的特征融合
netfourth = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu,
name='feature_fourth_bn')
net = denseblockpl(net, step=128, firstchannel=128, is_train=is_train, name='denseplz3', reuse=reuse)
# 第五次的特征融合
netfifth = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu,
name='feature_fifth_bn')
net = denseblockfin(net, step=128, firstchannel=128, is_train=is_train, name='denseplz4', reuse=reuse)
# 第六次的特征融合
netsixth = BatchNormLayer(net, is_train=is_train, decay=conv_bn_decay, act=tf.nn.relu,
name='feature_sixth_bn')
outfirst=Conv2d(netfirst, default_box_size[0] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='firstout')
outsecond=Conv2d(netsecond, default_box_size[1] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='secondout')
outthird=Conv2d(netthird, default_box_size[2] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='thirdout')
outfourth=Conv2d(netfourth, default_box_size[3] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='fourthout')
outfifth=Conv2d(netfifth, default_box_size[4] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='fifthout')
outsixth=Conv2d(netsixth, default_box_size[5] * (classes_size + 4), (3, 3), (1, 1), padding='SAME',
W_init=W_init, name='sixthout')
features1=outfirst.outputs
features2=outsecond.outputs
features3=outthird.outputs
features4=outfourth.outputs
features5=outfifth.outputs
features6=outsixth.outputs
feature_maps = [features1, features2, features3, features4, features5,features6]
global feature_maps_shape
feature_maps_shape = [m.get_shape().as_list() for m in feature_maps]
tmp_all_feature = []
for i, fmap in zip(range(len(feature_maps)), feature_maps):
width = feature_maps_shape[i][1]
height = feature_maps_shape[i][2]
tmp_all_feature.append(
tf.reshape(fmap, [-1, (width * height * default_box_size[i]), (classes_size + 4)]))
tmp_all_feature = tf.concat(tmp_all_feature, axis=1)
feature_class = tmp_all_feature[:, :, :classes_size]
feature_location = tmp_all_feature[:, :, classes_size:]
print('## feature_class shape : ' + str(feature_class.get_shape().as_list()))
print('## feature_location shape : ' + str(feature_location.get_shape().as_list()))
# 生成所有default boxs
global all_default_boxs
all_default_boxs = generate_all_default_boxs()
# print(all_default_boxs)
global all_default_boxs_len
all_default_boxs_len = len(all_default_boxs)
print('## all default boxs : ' + str(all_default_boxs_len))
return feature_class,feature_location,all_default_boxs,all_default_boxs_len
这里就是网络的主结构，这里netsecond-netsixth 好像并没有融合上提到的来自那个6层的输出，只是按照SSD进行了融合。
不过这里的代码比较简单对于做实验来说比较好修改，但我没调试。
希望有机会好好深入研究吧。。。。
参考：
DSOD: Learning Deeply Supervised Object Detectors from Scratch
网友评论