Training results cannot be reproduced about cmrnet

[cattaneod]

That's weird, I just tested again and my loss after one epoch is already lower than your results after 300:

epoch 0 total training loss = 1.332

total test loss = 1.208
total traslation error: 148.04042782727853 cm
total rotation error: 5.796896150946433 °

Can you share your environment, like pytorch version, cuda version, and so on?

[liyang159357]

The program runs on Windows platform,recompile "correlation_cuda","visibility" by vs2019.
In the code，RT = T@R instead of RT = T * R. Because blender has adjusted its mathutils module, replacing the asterisk * with the at symbol @
In the config,I rename ”kitti-00.csv“ to "poses.csv" in "./KITTI/sequences/00"

[caocheng1cc]

你可以告诉我怎么实现的吗，因为我没有找到pose.txt

[caocheng1cc]

In accordance with "python main_visibility_CALIB.py with batch_size=24 data_folder=./KITTI_ODOMETRY/ epochs=300 max_r=10 max_t=2 BASE_LEARNING_RATE=0.0001 savemodel=./checkpoints/ test_sequence=0"

My training results are as follows：

total test loss = 1.449
total traslation error: 164.9663721087535 cm
total rotation error: 6.641193625108441 °
This is 300-th epoch

Is there something wrong with the submitted code?

请问你的torch版本和torchvision版本是多少，我一直报rotat的错误

[Zhaohx424]

I use the pretrained model provided by the author and evaluate it in KITTI odometry datasets, but I got a very worse result than the results in the paper. For example, the median Transl.Error tested in sequence 00 is 19.2366cm, and even in the sequence 03 which used to train the model it still 10.6865cm, they are all much bigger than the results, I want to know why?
In addition, my environment is Ubuntu 18.04, cuda11.3, PyTorch1.10, Python3.6, they are all fit.

[Zhaohx424]

I use the pretrained model provided by the author and evaluate it in KITTI odometry datasets, but I got a very worse result than the results in the paper. For example, the median Transl.Error tested in sequence 00 is 19.2366cm, and even in the sequence 03 which used to train the model it still 10.6865cm, they are all much bigger than the results, I want to know why? In addition, my environment is Ubuntu 18.04, cuda11.3, PyTorch1.10, Python3.6, they are all fit.

OK I know, the results you put in Github uses cm, and the results in your paper uses m, I think you need to modify the error in Github readme which is wrong in system of unit.

[cattaneod]

I use the pretrained model provided by the author and evaluate it in KITTI odometry datasets, but I got a very worse result than the results in the paper. For example, the median Transl.Error tested in sequence 00 is 19.2366cm, and even in the sequence 03 which used to train the model it still 10.6865cm, they are all much bigger than the results, I want to know why? In addition, my environment is Ubuntu 18.04, cuda11.3, PyTorch1.10, Python3.6, they are all fit.

OK I know, the results you put in Github uses cm, and the results in your paper uses m, I think you need to modify the error in Github readme which is wrong in system of unit.

Hi @Zhaohx424,
you are right, the results in the readme use a wrong unit. thank you for pointing that out. I will correct it now.