guangmingw / doplearning Goto Github PK

there are three types of masks, what are their meaning?

I'd like to test the pretrained networks you provide on a custom, 82-frame dataset with just no scene changes. I looked at the code in stillbox_eval/depth_evaluation_utils.py, however much remains unclear. For instance:

• What is the code's understanding of scene, ref_imgs and tgt?
• Is the first frame of my sequence as the tgt and the rest sequentially listed in ref_imgs?
• What if I don't have any depth map to provide? I just want to test the pretrained dataset, not train the network myself

In short, could you provide a simple overview of the steps needed to load a custom dataset?

Hi! @guangmingw thanks for your codes.
I can't download the pretrained models, I gauss that it is because the URL cannot be accessed from the external network. Can you give a link that is publicly accessible？

Hi! Thanks a lot for releasing so useful code. Since it is an unsupervised method, how you deal with the scale problems in depth estimation? I mean that during testing how to ensure that the scale of estimated depth for every mono video be consistent?

Hi,

When running test_flow.py, I got the following error:

Traceback (most recent call last):
File "test_flow.py", line 396, in
main()
File "test_flow.py", line 160, in main
rigidity_mask_fwd = torch.zeros_like(rigidity_mask_fwd)
TypeError: zeros_like(): argument 'input' (position 1) must be Tensor, not list

Is there any way to fix the error? Thanks.

hello, after training posenet and dispnet by the command in readme. how to train the flownet?

there are 7 losses in train.py but the command in readme just used w1 and w3, how to use other losses, is their performance not good enough?

I run train.py according to the tutorial, but the following error is reported：
Traceback (most recent call last): File "/home/lcj/lcj/SLAM/DOPlearning-master/train.py", line 21, in <module> from loss_functions_summary import compute_errors, compute_epe, compute_all_epes, flow_diff, spatial_normalize_max, spatial_normalize_mean, new_static_mask, build_flow_not_intersect ImportError: cannot import name 'build_flow_not_intersect' from 'loss_functions_summary' (/home/lcj/lcj/SLAM/DOPlearning-master/loss_functions_summary.py)
I checked loss_functions_summary.py, there is no build_flow_not_intersect function here.

Hello, thanks for your codes.
I used your pretrained model for test_pose. but it does not performed well.

I use evo to evaluate the trajectory. the "gt" is sample['poses'] in test_pose.py line 91, and the "predictions" is your output. (cause both they are relative transform, I manually converted them into absolute transform.)

So I train the model according to your tutorial. But it goes worse after 2000 epochs.
I do not know why. Do you have some output of the posenet? Or can you give me some advise.

你好，感谢你分享代码。
我用你提供的model跑test_pose，但是画出来的轨迹表现很不好。（我在kitti_odometry的各个sequences上都试了下）
我使用evo工具来绘制轨迹。由于代码的输出是相对变换，我手动将test_pose.py第91行的predictions 和sample['poses']转换成绝对变换来绘制，他们的变化方式相同，因此应该不是这里有问题。
然后我打算自己训练，按照你提供的参数，只修改了路径。但是即使2000 epochs之后，跑出来的轨迹还是有问题。
我不知道是不是我的操作有问题，能否看一下你的posenet输出的可视化结果，又或者你能给我一些别的帮助。