Xingang Pan

@ezhonghawke Thanks for your interest. Unfortunately, these values are not available. Sorry about this inconvenience.

@voldemortX Thanks for your efforts. Your pytorch-auto-drive code base looks great! It's glad to know that your reproduction obtains even better results. I have added the link in `README.md`.

In our model, there is a branch predicting whether the lane lines exist. For no line scenes, that branch should predict that there is no lines.

@yoga-0125 Your understanding is correct. We believe that predicting the length accurately is also important for lane detection, thus length is considered in the evaluation metric.

@china56321 It seems that your lua torch is not successfully installed. Please follow the installation steps at https://github.com/facebookarchive/fb.resnet.torch/blob/master/INSTALL.md

@BoSeal You would affect the strength of your experiments. You may take a look at the tusimple dataset: https://github.com/TuSimple/tusimple-benchmark/issues/3

@BoSeal You can refer to https://github.com/XingangPan/SCNN-prototxt-generator for semantic segmentation version of SCNN.

@BoSeal Reducing batchsize would harm the performance of BN. One possible solution to this is to use frozen BN and smaller batchsize (like 2) as done in some semantic segmentation...

@OrkunYilmaz Err1 is semantic segmentation loss (spatial cross entropy), Err2 is classification loss (cross entropy) -backWeight: loss ratio for the background -maxIter: maximum number of iterations during training -retrain: path...

For cases where there are no lane markings in the image (i.e. crossroad), the groundtruth labeling is empty.