Questions about running the Cityscale dataset?

[116011]

I am very fortunate to have the opportunity to study your work. As for the SpaceNet dataset, I have successfully replicated. However, I encountered three issue with the Cityscale dataset.

1. when I followed the parameter file “toponet_vitb_512_cityscale.yaml” for reproduction, the inference of road centerlines for each test image fails completely. The issue manifests as the graph_points variable being empty around line 113 in the inferencer.py file. Could you please advise on how you resolved this?

2.during training, when using the function "def getitem (self, idx):" to take each sample, even if a random croppint is performed on the 2048x2048 sample, the data obtained in one epoch is 150 512x512 samples. At this time, training for 10 epochs, will there be insufficient training on the cityscale datase?

3. In some batches, when performing forward propagation through the model.py file at line 140 with self.transformer_encoder(), the shape of pair_features changes after it enters. This is contrary to what is described in your paper, where it is expected to remain unchanged. Could you clarify this?

Could you kindly provide the answer,thank you very much.

[EchoQiHeng]

I am very fortunate to have the opportunity to study your work. As for the SpaceNet dataset, I have successfully replicated. However, I encountered three issue with the Cityscale dataset.

1. when I followed the parameter file “toponet_vitb_512_cityscale.yaml” for reproduction, the inference of road centerlines for each test image fails completely. The issue manifests as the graph_points variable being empty around line 113 in the inferencer.py file. Could you please advise on how you resolved this?

2.during training, when using the function "def getitem (self, idx):" to take each sample, even if a random croppint is performed on the 2048x2048 sample, the data obtained in one epoch is 150 512x512 samples. At this time, training for 10 epochs, will there be insufficient training on the cityscale datase?

3. In some batches, when performing forward propagation through the model.py file at line 140 with self.transformer_encoder(), the shape of pair_features changes after it enters. This is contrary to what is described in your paper, where it is expected to remain unchanged. Could you clarify this?

Could you kindly provide the answer,thank you very much.

I don't quite understand "shape of pair_features changes."

[116011]

您好，我注意到您应该会说中文。上面提到的问题已经基本解决。我想另外请教一个问题，当你在从sapcenet数据集训练RNGDet++时，在main_sample.py(生成训练数据）、main_train.py(训练）中ROI_SIZE和image_size值给了多大？同时总训练轮次epochs设了多少？我在训练时，topo值最大为0.71（您训练的可以达到0.82），很难再上去。您方便加下微信吗，我的微信号是15927603180，将倍感荣幸。 武汉大学 王泽矫 2024-09-19 21:26:49 "EchoQiHeng" ***@***.***> 写道： I am very fortunate to have the opportunity to study your work. As for the SpaceNet dataset, I have successfully replicated. However, I encountered three issue with the Cityscale dataset. when I followed the parameter file “toponet_vitb_512_cityscale.yaml” for reproduction, the inference of road centerlines for each test image fails completely. The issue manifests as the graph_points variable being empty around line 113 in the inferencer.py file. Could you please advise on how you resolved this? 2.during training, when using the function "def getitem (self, idx):" to take each sample, even if a random croppint is performed on the 2048x2048 sample, the data obtained in one epoch is 150 512x512 samples. At this time, training for 10 epochs, will there be insufficient training on the cityscale datase? In some batches, when performing forward propagation through the model.py file at line 140 with self.transformer_encoder(), the shape of pair_features changes after it enters. This is contrary to what is described in your paper, where it is expected to remain unchanged. Could you clarify this? Could you kindly provide the answer,thank you very much. I don't quite understand "shape of pair_features changes." — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[htcr]

您好，感谢您的关注！ 我们没有自己训练RNGDet++，因为测试集和测试标准都是一样的，就直接引用原文的数据了。建议您和RNGDet++的作者直接联系咨询，如果不能复现我觉得训出来多少就写多少应该问题也不大...

[116011]

非常感谢您的回复，那City-scale数据集您用RNGDet++训练了吗。如果有的话，可以说下在main_sample.py(生成训练数据）、main_train.py(训练）中ROI_SIZE和image_size值给了多大？同时总训练轮次epochs设了多少？ 2024-09-24 13:31:32 "htcr" ***@***.***> 写道： 您好，感谢您的关注！ 我们没有自己训练RNGDet++，因为测试集和测试标准都是一样的，就直接引用原文的数据了。建议您和RNGDet++的作者直接联系咨询，如果不能复现我觉得训出来多少就写多少应该问题也不大... — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>