How to reproduced the results on Cityscapes?

[1225735136]

We reproduced the experiments on cityscapes, but get the different results. Is anything wrong with our operation?
First, we follow the SfMLearner wich resized the image to the size of --img_width=416 --img_height=171 and cropped the image to the size of --img_width=416 --img_height=128.
Second, we follow the instruction in your readme instruction which use the instruction as:
python manydepth2.train.py
--data_path
/home/jsw/datasets/cityscapes/cs_raw_sequence_format
--log_dir
logs
--png
--freeze_teacher_epoch
5
--model_name
models_many2_new10
--pytorch_random_seed
1
--batch_size
12
--mode
many2
--split
cityscapes_preprocessed
--dataset
cityscapes_preprocessed
--eval_split
cityscapes
--height
192
--width
512
Is there anything wrong with my instruction or operation? Help!

[kaichen-z]

Thank you for your interest. Please refer to ManyDepth for reproducing the Cityscapes results. In addition, we recommend experimenting with different values of freeze_teacher_epoch.

[1225735136]

Thank you for your reply! I tried to refer to the ManyDepth and SFMlearner. Following them, we transform the image size at 384*1024. Then we use the instruction similar as your instruction which is as follow:
python manydepth2.train.py
--data_path
/home/jsw/datasets/cityscapes/cs_raw_sequence_format
--log_dir
logs
--png
--freeze_teacher_epoch
5
--model_name
models_many2_new10
--pytorch_random_seed
1
--batch_size
12
--mode
many2
--split
cityscapes_preprocessed
--dataset
cityscapes_preprocessed
--eval_split
cityscapes
--height
192
--width
512
But we still can't reproduced the results. We Tried the trainning only on the cityscapes and got the a1 at 0.876. We tried to use the weight trained on kitti and finetuned on cityscapes and got the a1 at 0.874. Do we need to freeze some part of the student network or we neglect something else? Thank you for taking the time to reply to my message amid your busy schedule.

[hongaijie]

Hello there, I’m actually facing the same issue as well. I was wondering if you’ve been able to resolve it already? I’d really appreciate hearing back from you whenever you have a moment. Thank you so much!

[1225735136]

I write the results which i reproduced.

[kaichen-z]

Thank you for using our repository. Due to limited bandwidth, we are no longer actively maintaining or updating this repository. However, using the environment specified in the repository should allow you to reproduce the same results reported in our paper.

[Asuna88]

I write the results which i reproduced.

请问您最后是采用怎样的方式进行预处理和训练？
比如：

是把数据预处理到 384x1024尺寸吗？还是 处理到 416 x 171 尺寸？
训练的 --freeze_teacher_epoch 设置为几？
以及 在cityscapes从头训练的？还是 从kitti数据集训练好的模型再到cityscapes进行微调的？

谢谢您