网络结构

[haicaoha]

您好，想问一下当选择bisenetV2作为网络结构时，encoder部分是哪些，decoder部分又是哪些？两个分支就是共用的哪些部分？下边是这里建造模型的代码 def build_model(self, input_tensor, name, reuse=False): """

    :param input_tensor:
    :param name:
    :param reuse:
    :return:
    """
    with tf.variable_scope(name_or_scope=name, reuse=reuse):
        # build detail branch
        detail_branch_output = self.build_detail_branch(
            input_tensor=input_tensor,
            name='detail_branch'
        )
        # build semantic branch
        semantic_branch_output, _ = self.build_semantic_branch(
            input_tensor=input_tensor,
            name='semantic_branch',
            prepare_data_for_booster=False
        )
        # build aggregation branch
        aggregation_branch_output = self.build_aggregation_branch(
            detail_output=detail_branch_output,
            semantic_output=semantic_branch_output,
            name='aggregation_branch'
        )
        # build binary and instance segmentation branch
        binary_seg_branch_output = self.build_binary_segmentation_branch(
            input_tensor=aggregation_branch_output,
            name='binary_segmentation_branch'
        )
        instance_seg_branch_output = self.build_instance_segmentation_branch(
            input_tensor=aggregation_branch_output,
            name='instance_segmentation_branch'
        )
        # gather frontend output result
        self._net_intermediate_results['binary_segment_logits'] = {
            'data': binary_seg_branch_output,
            'shape': binary_seg_branch_output.get_shape().as_list()
        }
        self._net_intermediate_results['instance_segment_logits'] = {
            'data': instance_seg_branch_output,
            'shape': instance_seg_branch_output.get_shape().as_list()
        }
    return self._net_intermediate_results

[MaybeShewill-CV]

@haicaoha 这个你可以具体看一下原论文。然后对照着注释，注释里的分支名称基本都是按照论文来起的:)

[haicaoha]

请问是basenetV2的论文吗，我现在的主要问题是在lanenet论文里是编码-解码的网络结构，并且是两个分支网络；但是basenetV2的网络好像并不是编码-解码结构，并且好像二值分割和实例分割旨在最后网络尾部是不一样的，和lanenet论文中提到似乎不太一样，请问为什么这样设计网络呢

------------------ 原始邮件 ------------------ 发件人: "MaybeShewill-CV/lanenet-lane-detection" @.>; 发送时间: 2022年4月25日(星期一) 晚上6:20 @.>; @.**@.>; 主题: Re: [MaybeShewill-CV/lanenet-lane-detection] 网络结构 (Issue #553)

@haicaoha 这个你可以具体看一下原论文。然后对照着注释，注释里的分支名称基本都是按照论文来起的:)

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[MaybeShewill-CV]

@haicaoha lanenet的网络模型也是可以修改的 binary seg和instance seg只要共享特征提取就可以了，至于特征提取你用什么网络结构都是可以实验测试的。我这里为了效率高就用了轻量网络bisenetv2