Great work on your project! However, I cannot find the txt folder under the VIL100 dataset, which is used in your program(/RVLD/VIL100/txt/anno_txt/). Where should I find this folder or generate it?

感谢您优秀的工作。
在代码中我看到self.model.classifier.eval()，请问处于什么考量不训练classifier呢？

    def finetune_model(self):
        val1 = True
        val2 = False

        for param in self.model.regressor.parameters():
            param.requires_grad = val1  #
        for param in self.model.offset_regression.parameters():
            param.requires_grad = val1
        for param in self.model.deform_conv2d.parameters():
            param.requires_grad = val1
        for param in self.model.classifier.parameters():
            param.requires_grad = val2
        if val1 == False:
            self.model.regressor.eval() 
            self.model.offset_regression.eval()
            self.model.deform_conv2d.eval()
        if val2 == False:
            self.model.classifier.eval()

Great work! I would like to confirm whether the annotation of OpenLane-V is only for OpenLane's 2D lanes or also includes 3D lanes? Looking forward to your reply.

Thanks for your perfect work, I would like to know your training configuration for CLRNet, GANet on openlane, mainly learning rate, batch_size and epochs.

How do you test on CLRNet

Thanks for your work! However, I want to ask whether the pretrained model you provided can be directly used for reproduce your performance in your paper?

I've noticed that the video gets split when the vehicle changes lanes, which doesn't seem to make much sense .

Thank you for your excellent work and the provided dataset.

I have a question regarding the training process of CLRNet and GANet on the OpenLane-V dataset.
I attempted to reproduce the results using your dataset, but my scores did not align with those reported in the paper.
Could you provide assistance, possibly in the form of a config file or specific settings, for training these networks?

Thank you.

Great work on your project! However, in this paper, coefficient map is only generated from probability map. I want to know if using feature map as an additional input will result in a more accurate coefficient map? Have you conducted any relevant experiments?

Your paper states that some of the details are in the supplementary document, but I didn’t see it.

Great work on your project! I want to know when you plan to release the new dataset "OpenLane-V".

When I evaluated the authors' provided ILD_Coeff model, I did not get the F1 score of 0.787 reported in table 4 of the paper. Instead I got F1 score of 0.8056.

This is the exact output of the eval:

precision 0.7053766470275624, recall 0.5420338264165887, fscore 0.6130106817959742

culane laneatt metric evaluation start!

Loading prediction data...
100%|████████████████████████████████████████████████████████████████████████| 23021/23021 [00:08<00:00, 2827.61it/s]
Loading annotation data...
100%|████████████████████████████████████████████████████████████████████████| 23021/23021 [00:09<00:00, 2302.34it/s]
Calculating metric in parallel...
100%|█████████████████████████████████████████████████████████████████████████| 23021/23021 [02:50<00:00, 134.98it/s]
====================Results (datalist.txt)====================
TP: 62200
FP: 13849
FN: 16172
Precision: 0.8179
Recall: 0.7937
F1: 0.8056

culane laneatt metric evaluation done!

But I do get the authors' reported F1 score when I evaluate their provided PLD model, with an F1 score of 0.8249 which matches the paper's number in table 4 of 0.825.

This is the exact output of the eval for PLD:
precision 0.6302282598102822, recall 0.5900906007422002, fscore 0.6094992939678525

culane laneatt metric evaluation start!

Loading prediction data...
100%|█████████████████████████████████████████| 23021/23021 [00:07<00:00, 2891.16it/s]
Loading annotation data...
100%|█████████████████████████████████████████| 23021/23021 [00:10<00:00, 2285.51it/s]
Calculating metric in parallel...
100%|██████████████████████████████████████████| 23021/23021 [02:47<00:00, 137.51it/s]
miou ===> 0.7256893668357335
====================Results (datalist.txt)====================
TP: 64516
FP: 13541
FN: 13856
Precision: 0.8265
Recall: 0.8232
F1: 0.8249

culane laneatt metric evaluation done!

I am wondering if this is expected or if I did something wrong? Because this is not as much of an improvement from single-frame to video-based detection as is reported in the paper.

Thanks.

I would like to train a model which achieves your SOTA result of 0.825 [email protected] on OpenLane-V dataset. Which model/folder should I use for doing this? I assume I should use this one Modeling/OpenLane-V/PLD/code? Thank you for the amazing work.

Edit: Never mind. I should have more carefully read the paper. It seems you train ILD first and then train PLD. It also seems that you train ILD in two separate stages as well, the probability map stage and the coefficient map stage, so the model requires training 3 stages in total? Did you try training the model end-to-end? Was there a significant benefit to training these three stages separately? Thank you.

I am curious to see the generalization of model on different dataset, and to do this, wanted to visualize the output of your model on unseen image.

Hello, I would like to ask which GPU you are using. I use the 3090, but it takes a long time to train the openlane. and in addition, is there any multi-gpu training code.

Thanks for your contribution on new dataset.
But i have a question:
The label folder /OpenLane-V/label/training/ contains 450 clips, each containing a maximum of 199 images and a minimum of 5 images,
But the file /OpenLane-V/list/datalist_video_training.pickle contain 622 clips, each containing a maximum of 199 images and a minimum of 1 images. And both folders contain the same total of 68143 images.
Why split 450 clips into 622？

Hello, thanks for your great work. I have some questions about the function continuous_cross_iou. In the function continuous_cross_iou, the following code is used: image = Polygon([(0, 0), (0, h - 1), (w - 1, h - 1), (w - 1, 0)]). What is the effect of this line? Can I treat this line as the concept of optical flow?