cfzd
cfzd
@liujiachang If I understand you correctly, your problem is that you only have fixed-angle images for training while you need to test on the various angle images? I think you...
@Aruen24 Please install CULane evaluation tools as the `install.md` says.
@mimi37 https://github.com/TuSimple/tusimple-benchmark/issues/3
@dragonhaha Yes, you can refer to the model of CurveLanes. It could support 10 lanes.
@myasser63 You can directly add the FCA layer without any modification. The feature map's size would be addressed automatically as here: https://github.com/cfzd/FcaNet/blob/aa5fb63505575bb4e4e094613565379c3f6ada33/model/layer.py#L54-L55
Whatever you want. You can set it according to your preferences or use the settings as ours: https://github.com/cfzd/FcaNet/blob/aa5fb63505575bb4e4e094613565379c3f6ada33/model/fcanet.py#L19 https://github.com/cfzd/FcaNet/blob/aa5fb63505575bb4e4e094613565379c3f6ada33/model/fcanet.py#L29
@myasser63 As the error says, it's a problem of `adaptive_avg_pool2d`. You can either just ignore it by: ``` torch.use_deterministic_algorithms(True, warn_only=True) ``` or you can turn off determinism by: ``` torch.use_deterministic_algorithms(False)...
@yqytsq 掩码的GT只会用来训练分割分支（如果开启了的话），主体分类网络的训练主要是依赖json文件中的GT
@Li-Hanzhao Good work! @kidcats 可能是与训练数据集差异太大了
@yqytsq 这是一个非常好的问题。其实就是如何表示GT中的车道线的问题，我曾经尝试过使用参数曲线、分段函数等的表达方法，但最终还是受限于工程问题（dali需要每个batch数据等长），使用了行锚的形式表达。但我觉得行锚的使用并不会产生特别大的问题，这主要是因为GT的结果可以认为是没有误差的（相对于网络预测），所以锚上的误差放大效应在GT上可以认为影响不大，至少不会大到影响网络学习。确实可能会出现极端情况行锚无法描述的GT，但一般操作时会把这些GT给丢掉，即使丢掉少数GT还是不太影响最终结果。