Yifei Huang

> 请问bottom_priors_nums和left_priors_nums是指什么？ 初始的时候prior的分配数量，起点在图像左右边界各1/4，图像下边界1/2

Seed default is 0, you can set this in command line. https://github.com/Turoad/CLRNet/blob/51e082db12973943bddefd76fd0d431fcb3350ff/main.py#L68

> > Seed default is 0, you can set this in command line. > > https://github.com/Turoad/CLRNet/blob/51e082db12973943bddefd76fd0d431fcb3350ff/main.py#L68 > > Which seeds are used in your experience? use default value 0.

> `cd $CLRNET_ROOT/tools/lane_evaluation`, but the dir is not exists. Yes, this is a mistake, you can ignore this step, we have reimplemented the eval metrics using python opencv.

Do you mean the backbone? We use resnet18, resnet34, resnet101 and DLA34 to serve as our backbone and do following experiments.

> Hi! Thank you so much for this valuable work for the community. I would like to ask, in line 73 of clrnet/model/head/clrhead.py, is the role of self.cls_modules to judge...

This is a mistake in our code, this method is never used yet in `detector`, so it's ok to ignore this. We have moved this method to `heads` now, you...

> hi, It's a great work. But when I understand the source code, I'm confused with the meaning of "target_yxtl[:, -1] -= (predictions_starts - target_starts)" in clr_head.py 389 line. Can...

就是在最顶层的feature map上做了三次迭代，这里是做对比实验用的。

>  好的。谢谢您的回复。还有个问题。就是这个公式。 1、关于这个Pt的公式。这个Refined P0是不是就是将输出G叠加到feature map例如L0上的特征，然后当做L1的先验输入？  2、对于Refined是不是其中包含了ROIGather这些？我在另一个issue看到了类似的解释。那实际上refined是一个大的结构包含了处理P0的输入将其映射到L0 feature map。然后对这个新的feature map进行resize、flatten得到图中标注的Xf和卷积、FC得到Xp特征，然后进行attention操作得到W，然后再与Xf做特征抽取，得到新的车道线特征输入给feature map L1。这部分即Refined P0。 总结就是，我的理解Refined structure包含了中间整个部分。期待您的回复，谢谢！ 1. 首先每个stage的输入P_t是指lane prior的参数，即x, y, theta值（这三个参数值就能决定一条直线）。然后ROIGather模块是一个特征聚合的模块，用来生成和enhance对应lane prior的feature。在ROIGather之后会通过lane feature生成参数的regression值，输入 + regression值得到refine之后的下一个stage的输入。 2. 是的，公式当中的R_t代表的就是refine的过程，从输入prior参数到得到refined prior参数过程。中间包含了提feature、ROIGather等过程。不过并不是特征输入给L1，而是得到的refined prior params输入到下一个level。这个你可以看下代码的实现就能理解了～