-
SSD的整体结构流程图：
-
具体代码部分：
- ssd_net 网络结构：
# SSD net definition
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
end_points={}
with tf.variable_scope(scope,'ssd_300_vgg',[inputs],reuse=reuse):
# original vgg-16 block,if you want change base-line network,you can change in here
net = slim.repeat(inputs,2,slim.conv2d,64,[3,3],scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# block2
net = slim.repeat(net,2,slim.conv2d,128,[3,3],scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net,[2,2],scope='pool2')
# block3
net = slim.repeat(net,3,slim.conv2d,256,[3,3],scope='conv3')
end_points['block3']=net
net = slim.max_pool2d(net,[2,2],scope='pool3')
# block4
net = slim.repeat(net,3,slim.conv2d,512,[3,3],scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net,[2,2],scope='pool4')
# block5
net = slim.repeat(net,3,slim.conv2d,512,[3,3],scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net,[3,3],stride=1,scope='pool5')
# additional ssd blocks
# block6
net = slim.conv2d(net,1024,[3,3],rate=6,scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,rate=dropout_keep_prob,training=is_training)
# block 7:1x1 conv
net = slim.conv2d(net,1024,[1,1],scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,rate=dropout_keep_prob,training=is_training)
# block 8,9,10,11
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net,256,[1,1],scope='conv1x1')
net = custom_layers.pad2d(net,pad=(1,1))
net = slim.conv2d(net,512,[3,3],stride=2,scope='conv3x3',padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net,128,[1,1],scope='conv1x1')
net = custom_layers.pad2d(net,pad=(1,1))
net = slim.conv2d(net,256,[3,3],stride = 2,scope='conv3x3',padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net,128,[1,1],scope='conv1x1')
net = slim.conv2d(net,256,[3,3],scope='con3x3',padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net,128,[1,1],scope='conv1x1')
net = slim.conv2d(net,256,[3,3],scope='conv3x3',padding='VALID')
end_points[end_point] = net
#prediction and localisations layers
predictions =[]
logits = []
localisations = []
for i,layer in enumerate(feat_layers):
# 做多尺度大小box预测的特征层，返回每个cell中每个先验框的类别p和预测的位置l
p,l = ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions,localisations,logits,end_points
- 先验框生成：
def ssd_anchors_all_layers(img_shape,
layers_shape,
anchor_sizes,
anchor_ratios,
anchor_steps,
offset=0.5,
dtype=np.float32):
"""
计算所有特征图的先验框
:param img_shape: 输入图像shape
:param layers_shape: 特征图shape
:param anchor_sizes: 先验框大小
:param anchor_ratios:
:param anchor_steps: 先验框所在特征图相对于输入图像的比例
:param offset:
:param dtype:
:return:
"""
layers_anchors =[]
for i,s in enumerate(layers_shape):
anchor_bboxes = ssd_anchor_one_layer(img_shape,s,
anchor_sizes[i],
anchor_ratios[i],
anchor_steps[i],
offset=offset,
dtype = dtype)
layers_anchors.append(anchor_bboxes)
return layers_anchors
def ssd_anchor_one_layer(img_shape,
feat_shape,
sizes,
ratios,
step,
offset=0.5,
dtype=np.float32):
"""
compute SSD default anchor boxes for one feature layer.
determine the relative position grid of the centers, and the relative width and height
:param img_shape:
:param feat_shape:
:param sizes:
:param ratios:
:param step:
:param offset: grid offser
:param dtype:
:return: y,x,h,w:relative x and y grids,and height and width
"""
y,x = np.mgrid[0:feat_shape[0],0:feat_shape[1]]
y = (y.astype(dtype) + offset) * step /img_shape[0]
x = (x.astype(dtype)+offset) * step / img_shape[1]
# expanded dims to support easy broadcasting
x = np.expand_dims(x,axis=-1)
y = np.expand_dims(y, axis=-1)
# try to compute the relative height and width
num_anchors = len(sizes) + len(ratios)
h = np.zeros((num_anchors,), dtype=dtype)
w = np.zeros((num_anchors,),dtype=dtype)
h[0] = sizes[0] / img_shape[0]
w[0] = sizes[0] / img_shape[1]
di = 1
if len(sizes) > 1:
h[1] = math.sqrt(sizes[0] * sizes[1]) / img_shape[0]
w[1] = math.sqrt(sizes[0] * sizes[1]) / img_shape[1]
di +=1
for i,r in enumerate(ratios):
h[i+di] = sizes[0] / img_shape[0] / math.sqrt(r)
w[i+di] = sizes[0] / img_shape[1] * math.sqrt(r)
return y,x,h,w
- 默认框（先验框）匹配策略，寻找与默认框IOU最大的GTbox：
def bbbox_encode(self, labels, bboxes, anchors, scope=None):
"""
该部分主要是默认框的匹配策略（
和原论文中的Matching strategy有些不同，该部分仅仅是寻找与默认框IOU最大的GTbox，
并没有通过阈值0.5去筛选正样本），
将每个默认框与ground truth box进行匹配，
寻找与之IOU（交并比）最大的ground truth box，
并计算每个默认框与之匹配的ground truth box的偏差（
矩形框中心坐标x、y方向偏移量，以及高h宽w的缩放比例）
:param labels: 是GT box对应的标签
:param bboxes:bboxes是GT box对应的坐标信息
:param anchors: 生成的默认框
:param scope:
:return:
"""
return ssd_common.tf_ssd_bboxes_encode(labels,bboxes,anchors,
self.params.num_classes,
self.params.no_annotation_label,
ignore_threshold=0.5,
prior_scaling=self.params.prior_scaling,
scope=scope)
def tf_ssd_bboxes_encode(labels,
bboxes,
anchors,
num_classes,
no_annotation_label,
ignore_threshold=0.5,
prior_scaling=[0.1,0.1,0.2,0.2],
dtype=tf.float32,
scope='ssd_bboxes_encode'):
"""
encode groundtruth labels and bounding boxes using SSD net anchors.
encode boxes for all feature layers.
:param labels: 1D tensor(int 64) containing groundtruth labels 真实标签
:param bboxes: Nx4 tensor (float) with bboxes relative coordinate 真实bbox
:param anchors: list of Numpy array with layer anchors 存放每一个预测层生成的默认框
:param num_classes:
:param no_annotation_label:
:param ignore_threshold: threshold for positive match with groundtruce bboxes
:param prior_scaling: scaling of encoded coordinates
:param dtype:
:param scope:
:return:
(target_labels, target_localizations, target_scores):
each element is a list of target tensors
"""
with tf.name_scope(scope):
target_labels = [] # 存放匹配到的GTbox的label的 容器
target_localizations =[] # 存放匹配到的GTbox的位置信息的容器
target_scores =[] # 存放默认框与匹配到的GTbox的IOU（交并比）
for i, anchors_layer in enumerate(anchors): # 遍历每个预测层的默认框
with tf.name_scope('bboxes_encode_block_%i' % i):
t_labels, t_loc, t_scores = tf_ssd_bboxes_encode_layer(labels,bboxes, #匹配默认框的ground truth box并计算偏差
anchors_layer,
num_classes,
no_annotation_label,
ignore_threshold,
prior_scaling,dtype)
target_labels.append(t_labels) # 匹配到的ground truth box对应标签
target_localizations.append(t_loc) # 默认框与匹配到的ground truth box的坐标差异
target_scores.append(t_scores) # 默认框与匹配到的ground truth box的IOU（交并比）
return target_labels, target_localizations,target_scores
def tf_ssd_bboxes_encode_layer(labels, # GTbox类别
bboxes, # GTbox位置信息
anchors_layer, # 默认框坐标信息（中心点坐标以及宽/高
num_classes,
no_annotation_label,
ignore_threshold=0.5,
prior_scaling=[0.1,0.1,0.2,0.2],
dtype=tf.float32):
"""
在该函数中仅仅只是寻找与每个默认框最匹配的GTbox，
并没有进行筛选正负样本，关于正负样本的选取会在下一部分losses计算中讲述
encode groundtruth labels and bounding boxes using SSD anchors from one layer
:param labels: 1D Tensor(int 64) containing groundtruth labels
:param bboxes: Nx4 tensor (float) with bboxes relative coordinate
:param anchor_layer: numpy array with layer anchors
:param num_classes:
:param no_annotation_label:
:param ignore_threshold: threshold for positive match with groundtruth bboxes
:param prior_scaling: scaling of encoded coordinate
:param dtype:
:return:
(target_labels,target_localizations,target_scores): target tensors
"""
# anchors coordinate and volume 先验框的坐标和大小
# 1.得到每个anchor的（左上，右下）坐标,
# 因为 groundtruth_bboxes 的坐标表示为（y1,x1,y2,x2），
# 所以这一步将 anchor 的坐标也转换成这种形式。
# 转换到默认框的左上角以及右下角坐标
yref, xref, href, wref = anchors_layer
ymin = yref - href / 2.
xmin = xref - wref / 2.
ymax = yref + href / 2.
xmax = xref + wref / 2.
vol_anchors = (xmax - xmin) * (ymax - ymin) # 默认框的面积
# initialize tensors 初始化各种参数
# shape = （feat_size,feat_size,num_anchors）
shape = (yref.shape[0], yref.shape[1], href.size)
feat_labels = tf.zeros(shape, dtype=tf.int64) # 存放默认框匹配的GTbox标签
feat_scores = tf.zeros(shape, dtype=dtype) # 存放默认框与匹配的GTbox的IOU（交互比）
feat_ymin = tf.zeros(shape, dtype=dtype) # 存放默认框匹配到的GTbox的坐标信息
feat_xmin = tf.zeros(shape, dtype=dtype)
feat_ymax = tf.ones(shape, dtype=dtype)
feat_xmax = tf.ones(shape, dtype=dtype)
def jaccard_with_anchors(bbox): # 计算重叠度函数
"""
计算所有 anchors 和某一个 groundtruth_bbox 的IOU
compute jaccard score between a box and the anchors
:param bbox:
:return:
"""
int_ymin = tf.maximum(ymin, bbox[0])
int_xmin = tf.maximum(xmin, bbox[1])
int_ymax = tf.minimum(ymax, bbox[2])
int_xmax = tf.minimum(xmax, bbox[3])
h = tf.maximum(int_ymax - int_ymin, 0.)
w = tf.maximum(int_xmax - int_xmin, 0.)
# volumes
inter_vol = h * w
union_vol = vol_anchors - inter_vol + (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
jaccard = tf.div(inter_vol, union_vol)
return jaccard
def intersection_with_anchors(bbox):
"""
compute intersecion between score a box and the anchors
:param bbox:
:return:
"""
int_ymin = tf.maximum(ymin, bbox[0])
int_xmin = tf.maximum(xmin, bbox[1])
int_ymax = tf.minimum(ymax, bbox[2])
int_xmax = tf.minimum(xmax, bbox[3])
h = tf.maximum(int_ymax - int_ymin, 0.)
w = tf.maximum(int_xmax - int_xmin, 0.)
inter_vol = h * w
scores = tf.div(inter_vol, vol_anchors)
return scores
def condition(i, feat_labels, feat_scores, feat_ymin, feat_xmin, feat_ymax, feat_xmax):
"""
循环条件
condition: check label index
:return:
"""
r = tf.less(i, tf.shape(labels)) # tf.shape(labels)GTbox的个数，当i<=tf.shape(labels)时返回true
return r[0]
def body(i,feat_labels,feat_scores,feat_ymin,feat_xmin,feat_ymax,feat_xmax):
"""
执行循环主体
Body: update feature labels, scores and bboxes
follow the original SSD paper for that purpose:
- assign values when jaccard > 0.5
- only update if beat the score of other bboxes
寻找该层所有默认框匹配满足条件的GTbox
"""
# jaccard score
label = labels[i]
bbox = bboxes[i]
jaccard = jaccard_with_anchors(bbox) # 计算该层所有的默认框与该真实框的交并比
# mask : check threshold + scores + no annotations +num_classes
mask = tf.greater(jaccard, feat_scores) # 交并比是否比之前匹配的GTbox大
mask = tf.logical_and(mask, feat_scores > -0.5) # 暂时不清楚意义，但这里并不是为了获取正样本所以并不是大于0.5
mask = tf.logical_and(mask, label < num_classes) # 判断真实标签label小于num-classes，防止出错
imask = tf.cast(mask, tf.int64) # 转型
fmask = tf.cast(mask, dtype) # dtype float32
# update values using mask 根据mask更新标签和交并比
feat_labels = imask * label + (1 - imask) * feat_labels # 当imask为1时更新标签
feat_scores = tf.where(mask, jaccard, feat_scores) # 当mask为true时更新为jaccard，否则为feat-score
feat_ymin = fmask * bbox[0] + (1 - fmask) * feat_ymin # 当fmask为1.0时更新坐标信息
feat_xmin = fmask * bbox[1] + (1 - fmask) * feat_xmin
feat_ymax = fmask * bbox[2] + (1 - fmask) * feat_ymax
feat_xmax = fmask * bbox[3] + (1 - fmask) * feat_xmax
return [i+1, feat_labels, feat_scores,
feat_ymin, feat_xmin, feat_ymax, feat_xmax]
# main loop definition
i = 0
[i, feat_labels,feat_scores,
feat_ymin,feat_xmin,feat_ymax,feat_xmax] = tf.while_loop(condition,body, # condition是循环条件，body是循环体，第三项是参数
[i,feat_labels,feat_scores,
feat_ymin,feat_xmin,
feat_ymax,feat_xmax])
# transfer to center / size 转换回中心坐标以及宽高
feat_cy = (feat_ymax + feat_ymin) / 2.
feat_cx = (feat_xmax + feat_xmin) / 2.
feat_h = feat_ymax - feat_ymin
feat_w = feat_xmax - feat_xmin
# encode features 默认框中心与匹配的真实框中心坐标偏差，高和宽的偏差
feat_cy = (feat_cy - yref) / href / prior_scaling[0]
feat_cx = (feat_cx - xref) / wref / prior_scaling[1]
feat_h = tf.log(feat_h / href) / prior_scaling[2]
feat_w = tf.log(feat_w / wref) / prior_scaling[3]
# use SSD ordering: x / y/w/h instead of ours
feat_localizations = tf.stack([feat_cx,feat_cy,feat_w,feat_h],axis=-1)
return feat_labels,feat_localizations,feat_scores
- 预测框：
def detected_bboxes(self, predictions, localizations,
select_threshold=None, nms_threshold=0.5,
clipping_bbox=None, top_k=400, keep_top_k=200):
"""
通过SSD network输出得到预测框
:param predictions: 直接通过卷积得到的预测框种类
:param localizations: 直接通过卷积得到的预测框位置
:param select:
:return:
"""
# select top_k bboxes from preditions, and clip
# 得到对应某个类别的得分值以及bbox
rscores, rbboxes = ssd_common.tf_ssd_bboxes_select(predictions, localizations,
select_threshold=select_threshold,
num_classes=self.params.num_classes)
# 按照得分高低，筛选出400个bbox和对应得分
rscores, rbboxes = tfe.bboxes_sort(rscores, rbboxes, top_k=top_k)
# apply NMS algorithm 应用非极大值抑制，筛选掉与得分最高bbox重叠率大于0.5的，保留200个
rscores, rbboxes = tfe.bboxes_nms_batch(rscores, rbboxes,
nms_threshold=nms_threshold,
keep_top_k=keep_top_k)
if clipping_bbox is not None:
rbboxes = tfe.bboxes_clip(clipping_bbox, rbboxes)
return rscores, rbboxes
- 损失函数
def ssd_losses(logits, localisations, gclasses, glocalisations, gscores,
match_threshold=0.5,
negative_ratio=3.,
alpha=1.,
label_smoothing=0,
device='/cpu:0',
scope=None):
"""
:param logits: 预测类别
:param localisations: 预测偏移位置
:param gclasses: 正确类别
:param glocalisations: 实际偏移位置
:param gscores: 与GT的交并比
:param match_threshold:
:param negative_ratio:
:param alpha:
:param label_smoothing:
:param device:
:param scope:
:return:
"""
with tf.name_scope(scope,'ssd_losses'):
lshape = tfe.get_shape(logits[0],5)
num_classes = lshape[-1]
batch_size = lshape[0]
# flattern out all vectors 展平所有的向量
flogits = []
fgclasses = []
fgscores = []
flocalisations = []
fglocalisations = []
for i in range(len(logits)):
flogits.append(tf.reshape(logits[i], [-1, num_classes]))
fgclasses.append(tf.reshape(gclasses[i], [-1]))
fgscores.append(tf.reshape(gscores[i], [-1]))
flocalisations.append(tf.reshape(localisations[1], [-1, 4]))
fglocalisations.append(tf.reshape(glocalisations[i], [-1, 4]))
# add concat the crap
logits = tf.concat(flogits, axis=0)
gclasses = tf.concat(fgclasses, axis=0)
gscores = tf.concat(fgscores, axis=0)
localisations = tf.concat(flocalisations, axis=0)
glocalisations = tf.concat(fglocalisations, axis=0)
dtype = logits.dtype
# compute positive matching mask.. 计算正样本数目
pmask = gscores > match_threshold # 交并比是否大于0.5
fpmask = tf.cast(pmask, dtype)
n_positives = tf.reduce_sum(fpmask) # 正样本数目
# hard negative mining
no_classes = tf.cast(pmask, tf.int32)
predictions = slim.softmax(logits)
nmask = tf.logical_and(tf.logical_not(pmask), gscores > -0.5) # 交并比小于0.5并大于-0.5的负样本
fnmask = tf.cast(nmask, dtype) # 转成float型
nvalues = tf.where(nmask, predictions[:, 0], 1. - fnmask) # True时为背景概率， False时为1.0, 0是背景
nvalues_flat = tf.reshape(nvalues, [-1])
# number of negative entries to select
max_neg_entries = tf.cast(tf.reduce_sum(fnmask), tf.int32) # 所有供选择的负样本数目
n_neg = tf.cast(negative_ratio * n_positives, tf.int32) + batch_size
n_neg = tf.minimum(n_neg, max_neg_entries) # 负样本的个数
val, idxes = tf.nn.top_k(-nvalues_flat, k=n_neg) # 按排序获取前k个值，以及对应的id
max_hard_pred = -val[-1] # 负样本的背景概率阈值
# final negative mask
nmask = tf.logical_and(nmask, nvalues < max_hard_pred) # 交并比小于0.5并且大于-0.5的负样本，且概率小于max_hard_pred
fnmask = tf.cast(nmask, dtype)
# add cross-entropy loss
with tf.name_scope('cross_entropy_pos'):
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=gclasses)
loss = tf.div(tf.reduce_sum(loss * fpmask), batch_size, name='value') # fpmask是正样本的mask， 正为1, 负为0
tf.losses.add_loss(loss)
with tf.name_scope('cross_entropy_neg'):
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=no_classes)
loss =tf.div(tf.reduce_sum(loss * fnmask), batch_size, name='value') # fnmask是负样本的mask，负为1,正为0
tf.loss.add_loss(loss)
# add localization loss: smooth L1, L1, ...
with tf.name_scope('localization'):
# weights tensor: positive mask + random negative
weights = tf.expand_dims(alpha * fpmask, axis=-1)
loss = custom_layers.abs_smooth(localisations - glocalisations)
loss = tf.div(tf.reduce_sum(loss * weights), batch_size, name='value')
tf.losses.add_loss(loss)
参考大神代码来理解：大神666～
1042
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
