Segmentation branch
解决样本分布不均衡
车道线像素远小于背景像素.loss函数的设计对不同像素赋给不同权重,降低背景权重.
该分支的输出为(w,h,2)．
Embedding branch
loss的设计思路为使得属于同一条车道线的像素距离尽量小,属于不同车道线的像素距离尽可能大.即Discriminative loss.
该分支的输出为(w,h,n)．n为表示像素的向量的维度.
实例分割
在Segmentation branch完成语义分割,Embedding branch完成像素的向量表示后,做聚类,完成实例分割.
H-net
透视变换
to do
车道线拟合
LaneNet的输出是每条车道线的像素集合，还需要根据这些像素点回归出一条车道线。传统的做法是将图片投影到鸟瞰图中，然后使用二次或三次多项式进行拟合。在这种方法中，转换矩阵H只被计算一次，所有的图片使用的是相同的转换矩阵，这会导致坡度变化下的误差。
为了解决这个问题，论文训练了一个可以预测变换矩阵H的神经网络HNet，网络的输入是图片，输出是转置矩阵H。之前移植过Opencv逆透视变换矩阵的源码，里面转换矩阵需要8个参数，这儿只给了6个参数的自由度，一开始有些疑惑，后来仔细阅读paper，发现作者已经给出了解释，是为了对转换矩阵在水平方向上的变换进行约束。
代码分析
binary_seg_image, instance_seg_image = sess.run( [binary_seg_ret, instance_seg_ret], feed_dict={input_tensor: [image]} )输入(１,256,512,3)输出binary_seg_image:(1, 256, 512) instance_seg_image:(1, 256, 512, 4)
完成像素级别的分类和向量表示
class LaneNet的inference分为两步．
第一步提取分割的特征,包括了用于语义分割的特征和用以实例分割的特征.
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 ) #得到一个字典,包含了用于语义分割的feature map和用于实例分割的feature map． #binary_segment_logits (1,256,512,2) 2是类别数目.即车道/背景． #instance_segment_logits (1,256,512,64) 用以后面再做卷积为每个像素生成一个向量表示 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第一步得到的features如下:
features : OrderedDict([('encode_stage_1_share', {'data': , 'shape': [1, 256, 512, 64]}), ('encode_stage_2_share', {'data': , 'shape': [1, 128, 256, 128]}), ('encode_stage_3_share', {'data': , 'shape': [1, 64, 128, 256]}), ('encode_stage_4_share', {'data': , 'shape': [1, 32, 64, 512]}), ('encode_stage_5_binary', {'data': , 'shape': [1, 16, 32, 512]}), ('encode_stage_5_instance', {'data': , 'shape': [1, 16, 32, 512]}), ('binary_segment_logits', {'data': , 'shape': [1, 256, 512, 2]}), ('instance_segment_logits', {'data': , 'shape': [1, 256, 512, 64]})])特征提取完毕,做后处理
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对每个像素的分类,做softmax转成概率.再argmax求概率较大值的下标.对每个像素的向量表示,用1x1卷积核做卷积,得到channel维度=CFG.TRAIN.EMBEDDING_FEATS_DIMS(配置为4).即(1,256,512,64)卷积得到(1,256,512,4)的tensor．即每个像素用一个四维向量表示.
所以,整个LaneNet的inference返回的是两个tensor.一个shape为(1,256,512) 一个为(1,256,512,4).
后处理
class LaneNetPostProcessor(object): def postprocess(self, binary_seg_result, instance_seg_result=None, min_area_threshold=100, source_image=None, data_source='tusimple'):对binary_seg_result,先通过形态学操作将小的空洞去除.参考 https://www.cnblogs.com/sdu20112013/p/11672634.html
然后做聚类.
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获取到坐标及对应坐标像素对应的向量表示.
np.where(condition)
只有条件 (condition)，没有x和y，则输出满足条件 (即非0) 元素的坐标 (等价于numpy.nonzero)。这里的坐标以tuple的形式给出，通常原数组有多少维，输出的tuple中就包含几个数组，分别对应符合条件元素的各维坐标。
测试结果
tensorflow-gpu 1.15.2
4张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
类似于高吞吐量,高延迟.对单帧图片处理在1-2s,多幅图片同时处理,平均下来的处理速度在0.1s.
论文里的backbone为enet,在nvida 1080 ti上推理速度52fps.
对于这个问题的解释,作者的解释是
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:)
一个是特征提取网络和论文里不一致,一个是gpu有一个短暂的warm up的时间.
我自己的测试结果是在extract image features耗时较多.换一个backbone可能会有改善.
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
