Segmentation branch
解決樣本分佈不均衡
車道線像素遠小於背景像素.loss函數的設計對不同像素賦給不同權重,降低背景權重.
該分支的輸出為(w,h,2)．
Embedding branch
loss的設計思路為使得屬於同一條車道線的像素距離儘量小,屬於不同車道線的像素距離儘可能大.即Discriminative loss.
該分支的輸出為(w,h,n)．n為表示像素的向量的維度.
實例分割
在Segmentation branch完成語義分割,Embedding branch完成像素的向量表示後,做聚類,完成實例分割.
H-net
透視變換
to do
車道線擬合
LaneNet的輸出是每條車道線的像素集合，還需要根據這些像素點回歸出一條車道線。傳統的做法是將圖片投影到鳥瞰圖中，然後使用二次或三次多項式進行擬合。在這種方法中，轉換矩陣H只被計算一次，所有的圖片使用的是相同的轉換矩陣，這會導致坡度變化下的誤差。
為了解決這個問題，論文訓練了一個可以預測變換矩陣H的神經網絡HNet，網絡的輸入是圖片，輸出是轉置矩陣H。之前移植過Opencv逆透視變換矩陣的源碼，裡面轉換矩陣需要8個參數，這兒只給了6個參數的自由度，一開始有些疑惑，後來仔細閱讀paper，發現作者已經給出瞭解釋，是為了對轉換矩陣在水平方向上的變換進行約束。
代碼分析
<code>binary_seg_image, instance_seg_image = sess.run(
[binary_seg_ret, instance_seg_ret],
feed_dict={input_tensor: [image]}
)/<code>
輸入(１,256,512,3)輸出binary_seg_image:(1, 256, 512) instance_seg_image:(1, 256, 512, 4)
完成像素級別的分類和向量表示
class LaneNet的inference分為兩步．
第一步提取分割的特徵,包括了用於語義分割的特徵和用以實例分割的特徵.
<code>class LaneNet(cnn_basenet.CNNBaseModel):
def inference(self, input_tensor, name):
"""
:param input_tensor:
:param name:
:return:
"""
with tf.variable_scope(name_or_scope=name, reuse=self._reuse):
# first extract image features
extract_feats_result = self._frontend.build_model(
input_tensor=input_tensor,
name='{:s}_frontend'.format(self._net_flag),
reuse=self._reuse
)
#得到一個字典,包含了用於語義分割的feature map和用於實例分割的feature map．
#binary_segment_logits (1,256,512,2) 2是類別數目.即車道/背景．
#instance_segment_logits (1,256,512,64) 用以後面再做卷積為每個像素生成一個向量表示
print('features:',extract_feats_result)
# second apply backend process
binary_seg_prediction, instance_seg_prediction = self._backend.inference(
binary_seg_logits=extract_feats_result['binary_segment_logits']['data'],
instance_seg_logits=extract_feats_result['instance_segment_logits']['data'],
name='{:s}_backend'.format(self._net_flag),
reuse=self._reuse
)
if not self._reuse:
self._reuse = True
return binary_seg_prediction, instance_seg_prediction/<code>
第一步得到的features如下:
<code>features : OrderedDict([('encode_stage_1_share', {'data': <tf.tensor>, 'shape': [1, 256, 512, 64]}), ('encode_stage_2_share', {'data': <tf.tensor>, 'shape': [1, 128, 256, 128]}), ('encode_stage_3_share', {'data': <tf.tensor>, 'shape': [1, 64, 128, 256]}), ('encode_stage_4_share', {'data': <tf.tensor>, 'shape': [1, 32, 64, 512]}), ('encode_stage_5_binary', {'data': <tf.tensor>, 'shape': [1, 16, 32, 512]}), ('encode_stage_5_instance', {'data': <tf.tensor>, 'shape': [1, 16, 32, 512]}), ('binary_segment_logits', {'data': <tf.tensor>, 'shape': [1, 256, 512, 2]}), ('instance_segment_logits', {'data': <tf.tensor>, 'shape': [1, 256, 512, 64]})])/<tf.tensor>/<tf.tensor>/<tf.tensor>/<tf.tensor>/<tf.tensor>/<tf.tensor>/<tf.tensor>/<tf.tensor>/<code>特徵提取完畢,做後處理
<code>class LaneNetBackEnd(cnn_basenet.CNNBaseModel):
def inference(self, binary_seg_logits, instance_seg_logits, name, reuse):
"""
:param binary_seg_logits:
:param instance_seg_logits:
:param name:
:param reuse:
:return:
"""
with tf.variable_scope(name_or_scope=name, reuse=reuse):
with tf.variable_scope(name_or_scope='binary_seg'):
binary_seg_score = tf.nn.softmax(logits=binary_seg_logits)
binary_seg_prediction = tf.argmax(binary_seg_score, axis=-1)
with tf.variable_scope(name_or_scope='instance_seg'):
pix_bn = self.layerbn(
inputdata=instance_seg_logits, is_training=self._is_training, name='pix_bn')
pix_relu = self.relu(inputdata=pix_bn, name='pix_relu')
instance_seg_prediction = self.conv2d(
inputdata=pix_relu,
out_channel=CFG.TRAIN.EMBEDDING_FEATS_DIMS,
kernel_size=1,
use_bias=False,
name='pix_embedding_conv'
)
return binary_seg_prediction, instance_seg_prediction/<code>
對每個像素的分類,做softmax轉成概率.再argmax求概率較大值的下標.對每個像素的向量表示,用1x1卷積核做卷積,得到channel維度=CFG.TRAIN.EMBEDDING_FEATS_DIMS(配置為4).即(1,256,512,64)卷積得到(1,256,512,4)的tensor．即每個像素用一個四維向量表示.
所以,整個LaneNet的inference返回的是兩個tensor.一個shape為(1,256,512) 一個為(1,256,512,4).
後處理
<code>class LaneNetPostProcessor(object):
def postprocess(self, binary_seg_result, instance_seg_result=None,
min_area_threshold=100, source_image=None,
data_source='tusimple'):/<code>
對binary_seg_result,先通過形態學操作將小的空洞去除.參考 https://www.cnblogs.com/sdu20112013/p/11672634.html
然後做聚類.
<code>def _get_lane_embedding_feats(binary_seg_ret, instance_seg_ret):
"""
get lane embedding features according the binary seg result
:param binary_seg_ret:
:param instance_seg_ret:
:return:
"""
idx = np.where(binary_seg_ret == 255) #idx (b,h,w)
lane_embedding_feats = instance_seg_ret[idx]
# idx_scale = np.vstack((idx[0] / 256.0, idx[1] / 512.0)).transpose()
# lane_embedding_feats = np.hstack((lane_embedding_feats, idx_scale))
lane_coordinate = np.vstack((idx[1], idx[0])).transpose()
assert lane_embedding_feats.shape[0] == lane_coordinate.shape[0]
ret = {
'lane_embedding_feats': lane_embedding_feats,
'lane_coordinates': lane_coordinate
}
return ret/<code>
獲取到座標及對應座標像素對應的向量表示.
np.where(condition)
只有條件 (condition)，沒有x和y，則輸出滿足條件 (即非0) 元素的座標 (等價於numpy.nonzero)。這裡的座標以tuple的形式給出，通常原數組有多少維，輸出的tuple中就包含幾個數組，分別對應符合條件元素的各維座標。
測試結果
tensorflow-gpu 1.15.2
4張titan xp
(4, 256, 512) (4, 256, 512, 4)
I0302 17:04:31.276140 29376 test_lanenet.py:222] imgae inference cost time: 2.58794s
(32, 256, 512) (32, 256, 512, 4)
I0302 17:05:50.322593 29632 test_lanenet.py:222] imgae inference cost time: 4.31036s
類似於高吞吐量,高延遲.對單幀圖片處理在1-2s,多幅圖片同時處理,平均下來的處理速度在0.1s.
論文裡的backbone為enet,在nvida 1080 ti上推理速度52fps.
對於這個問題的解釋,作者的解釋是
2.Origin paper use Enet as backbone net but I use vgg16 as backbone net so speed will not get as fast as that. 3.Gpu need a short time to warm up and you can adjust your batch size to test the speed again:)
一個是特徵提取網絡和論文裡不一致,一個是gpu有一個短暫的warm up的時間.
我自己的測試結果是在extract image features耗時較多.換一個backbone可能會有改善.
<code>def inference(self, input_tensor, name):
"""
:param input_tensor:
:param name:
:return:
"""
print("***************,input_tensor shape:",input_tensor.shape)
with tf.variable_scope(name_or_scope=name, reuse=self._reuse):
t_start = time.time()
# first extract image features
extract_feats_result = self._frontend.build_model(
input_tensor=input_tensor,
name='{:s}_frontend'.format(self._net_flag),
reuse=self._reuse
)
t_cost = time.time() - t_start
glog.info('extract image features cost time: {:.5f}s'.format(t_cost))
# second apply backend process
t_start = time.time()
binary_seg_prediction, instance_seg_prediction = self._backend.inference(
binary_seg_logits=extract_feats_result['binary_segment_logits']['data'],
instance_seg_logits=extract_feats_result['instance_segment_logits']['data'],
name='{:s}_backend'.format(self._net_flag),
reuse=self._reuse
)
t_cost = time.time() - t_start
glog.info('backend process cost time: {:.5f}s'.format(t_cost))
if not self._reuse:
self._reuse = True
return binary_seg_prediction, instance_seg_prediction/<code>
閱讀更多 sandag 的文章
