The results of table 3 in paper

[fornote]

Hello, thanks for your outstanding work. Could you have open the evaluation metrics code of REG and REG Norm in table 3 of paper.

[cfzd]

@fornote Basically, the methods of REG and REG Norm are not in the scope of our work. They are just used to verify the effectiveness of our method. So, I would not provide the source code of these experiments. But I can tell you how we implemented these methods.

The big idea is very simple: converting a classification network to a regression network.

So you have to:

• modify the dimension of the network: https://github.com/cfzd/Ultra-Fast-Lane-Detection/blob/60f477c7358bbe177e1117b9f229a4a4b0db0e73/model/model.py#L63 make the total_dim equal to num_cls_per_lane x num_of_lanes, num_gridding is no longer needed.
• modify the data pipeline The original data pipeline is for classification, you need to modify it for regression. So, remove this line: https://github.com/cfzd/Ultra-Fast-Lane-Detection/blob/60f477c7358bbe177e1117b9f229a4a4b0db0e73/data/dataset.py#L81 Return the lane coordinate from lane_pts directly.
• modify the loss modify the losses in this function: https://github.com/cfzd/Ultra-Fast-Lane-Detection/blob/60f477c7358bbe177e1117b9f229a4a4b0db0e73/utils/factory.py#L28 make the classification loss to regression loss like the L1 loss.

Then, you can train the regression network. But if you want to test the trained regression network, you might also need to:

• modify the evaluation pipeline. modify the functions: https://github.com/cfzd/Ultra-Fast-Lane-Detection/blob/60f477c7358bbe177e1117b9f229a4a4b0db0e73/evaluation/eval_wrapper.py#L64 and https://github.com/cfzd/Ultra-Fast-Lane-Detection/blob/60f477c7358bbe177e1117b9f229a4a4b0db0e73/evaluation/eval_wrapper.py#L9

[Tianwei-She]

@cfzd Thanks for the detailed explanation! Just to confirm, the regression loss you used in REG and REG Norm experiments is L1 loss, right?

[fornote]

@cfzd Thank you very much. I have another question that if you evaluated the REG method with the same way of CLS?

[cfzd]

@Tianwei-She Yes, I tried both L1 and L2 loss. The L1 loss is more stable. I think smoothL1 loss could also work. @fornote Yes, they are evaluated in the same way. The only difference is the part that converts the outputs of the network to the evaluation format. For classification, you need to remap from the grid to coordinate. For regression, you may only need to rescale the output, since the output of regression is already the (scaled) coordinate. The evaluation is the same.

[ChenzrMax]

hello,

I have a question about Ultra-Fast Regression: how to express x_coordinate when lane don't exist?
and which activation function did you choose?
In my attempt, I express x = -1 when lane do not exist and choose Sigmoid function. And I got acc =92.98 on Tusimple Dataset. 

BR, Max

[cfzd]

@hitczr Once the issue is closed, I will no longer be informed of any new comments. Sorry for the very late response.

I used an additional classifier to indicate the case when the lane doesn't exist for each regression.

It seems your method works well with regression. When I was doing the ablation study, I also doubt the low performance of regression. But the results reported in the paper are already the best I can get. It would be very kind of you that describe the regression method. But I think the classification-based method would still have higher performance.

By the way, when you use Sigmoid, the output would range in [0,1], then how can the x equal -1?

[ChenzrMax]

@cfzd Thanks for your reply. As you said, when use sigmoid, x can not be equal to -1, it would just close to 0 when inference. I also tried activate function such as tanh that range in [-1, 1]. I agree with you that classification-based method is more effective.