A question about the test result

[khw321]

@MaybeShewill-CV Hi, I have trained on CULane dataset, and this is a strange test result, have you ever met this? And when two lanes are close, they are easy to become wider and be connected together, you can see it in the middle of the image.

[MaybeShewill-CV]

@khw321 That is caused by the mean shift cluster. The two closing lanes are clustered together so the polyline function draw them together.

[khw321]

@MaybeShewill-CV I got it, and the blue area in the right is caused by the cluster error. I have another question about changing the result format to Tusimple's format, My way is simply choosing the middle of the lane. What is yours strategy, and can you give me some advice on using the Hnet?

[MaybeShewill-CV]

@khw321 The HNet code has some bug for now and I will release it once it got fixed

[khw321]

@MaybeShewill-CV About the result format, what is your strategy to get the lane point?

[MaybeShewill-CV]

@khw321 The lane point is captured by the lanenet binary segmentation network. You can check the paper for details

[khw321]

@MaybeShewill-CV Sorry, my point is how to get such format result from the instance_seg_image. My way is simply choosing the middle of the lane according to the h_samles.

[MaybeShewill-CV]

@khw321 You can use the binary segmentation result as mask to get the instance lane points on the image.

[khw321]

@MaybeShewill-CV OK, thanks!

[cardwing]

@khw321 , since you mentioned that you have trained your model in CULane dataset, I wonder how well your model performs (e.g., F1 measure in different road categories and in total). I just want to know the performance of the model so that I can decide whether I should use this code. Thanks a lot ! ^_^

[khw321]

@cardwing I only used the images which have more than 1 lane in CULane dataset for training, the F1 measure in total is 63.2

[PenghuahuaPeng]

@khw321 Hi，i have trained lanenet model in CULane dataset,but in first evl epoch i got a "nan" loss. I have tried to change learning rate,batch_size and so on,which has turned out to be no use.So,do you think there is wrong with my traslated dataset?Can you show me your codes used for forming lanenet labels of culane dataset?Thank you very much.

[khw321]

@PenghuahuaPeng Sorry, I can't find my code now. You could check the label, the no lane label may cause error.

[PenghuahuaPeng]

Thanks a lot, i will have a try!

[PenghuahuaPeng]

@khw321 Sorry,excuse me,do you remember the final accuracy in culane dataset trained by these code?I have a 60-70 training accuracy and a 45 val accuracy after 60k epoch,now,it seems reach a balance,which is too slow.Thank you

[khw321]

@PenghuahuaPeng I only used the images which have more than 1 lane in CULane dataset for training, the F1 measure in total is 63.2

[PenghuahuaPeng]

thanks very much！

[abhigoku10]

@khw321 hi from the conversation above it is evident that u have trained a model on CU simple dataset is it possible for you to share the trained model

[TGLTommy]

@khw321 您好，请问如何将CULane的training数据集转换成TuSimple数据集格式？麻烦指点一下

[khw321]

@TommyTang930 CULane和Tusimple都是在固定高度上取点，只不过CULane的label是每根车道线一行，包含每个点的w,h坐标，而Tusimple是将h提出来作为h_sample，单独一行，然后每根车道线的w坐标为一行，如果某个h上不存在车道线的点，那对应的w设为-2

[khw321]

@abhigoku10 Sorry, I can't find my trained model now.