meteorshowers / x-stereolab Goto Github PK

“1.32 EPE_all with 8X single model 1.48EPE_all with 8X multi model on sceneflow dataset by end-to-end training.”
请问 8X single model 指的是1/8分辨率的未经过分层细化的模型吗？
8X multi model指的是1/8分辨率并且经过多层(3级)细化后的模型吗？
请问作者，经过细化后的模型的错误率(EPE)不是应该比不refine的要低吗？还是小白理解的有问题？请大神指点一二~~~
万分感谢~~~~~~~~

Is it possible you could you share some of the trained models you have created?

Is there a plan to release the source code on this project? I'd like to test-drive it myself.

Hello, thanks for your good implementation.
You deleted StereoNet code in this commit (e2cd703 ).
What a kind of bug you found?

你好，
仔细对照了你的代码和论文的原文，发现有个地方好像和论文不太对应。具体问题是：
在基于边缘的视差修正的模块里里面，我看到你的代先将小分辨的视差图进行了bilinear的上采样，然后又根据原始图像的宽度和小分辨视差图的宽度计算了比值，如果大于1.5就乘以系数8.这样做是有什么考虑么？我好像没在原文看到有这个操作。

谢谢啦

thanks for sharing the code, but the main8Xmulti.py script uses disparity gound truth, my question is : how to do the unsupervised training?

    for i in range(data.shape[0]):
        for j in range(data.shape[1]):
            if j - data[i][j] < 0:
                data[i][j] = -1

It make black areas at bottom-left corner of the disparity image.

When i try to train Sceneflow, this problem occur.

大神您好，我用GTX1060-6G显卡的笔记本跑您的模型，测得FPS=20；用Tesla K80-12G服务器跑您的模型测得FPS=8.，这两台设备跑的速度远达不到您所提到的30-25FPS以及达不到论文里的要求，不知道除了设备自身的问题还有哪些问题可以考虑？ 另外，用这个在SceneFlow上的预训练模型做KITTI数据集上的微调得到的效果很差，请问大神，您能否放出微调的部分代码供参考？非常感谢大神！

----再一次表示感谢！蟹蟹~~~

You can atleast release the first version of HITNet, i just want the same accuracy as mentioned in the paper even some offset will be also do fine,and the code need not be optimized.

I have tested the stereonet model pretrained from sceneflow dataset. However I got 2.93 epe which is much larger than 1.38. Should the pretrained model be finetuned in sceneflow dataset to reach 1.38?

欢迎加入深度学习双目视觉，群聊号码：1018698420

Particularly interested in activestereonet training.

I would like to ask the next author, if the size of the image used for network training is 960x540, then when I want to use this network test, the resolution of the photos taken by my camera is very large, such as 4608x3456. When I take this picture to train, the parallax map effect is very bad. When I reduce the original image to 960x540, the parallax map effect is very good OK, but I don't know how to restore the parallax value of the original image. Hope that the author can guide, thank you very much!

Hello, I can't find the "StereoNet" module in models.StereoNet8Xmulti

Hi The link for pretrain model is 404，Could you provide the pretrain model for StereoNet？ Thanks

Hi,
Thanks for your work!
May I ask which cuda version did you use?

I downloaded the stereo 2015/flow 2015/scene flow 2015 data set (2 GB) but I don't see any subfolder with the name 'dispnorm_occ'. Without it, I can't call disparity.KITTIloader2015.dataloader to actually train the StereoNet model. Can you help me find it or explain how to generate it?

Thanks!

Talk is cheap show me the code!!!

cannot convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first

Thank you for sharing the code.
Is it poosible for you to share the pre-trained model on sceneflow data?

How do I run the code of SOS?

Hi,
Thanks for sharing your code. I have confusion about the speed of the model.
As described in the paper，

Refer to the illustration in Table 2,
The StereoNet used a model with a downsampling factor of 8 and 3 refinement levels (i.e., StereoNet-8x-multi) can reach a runtime=0.015s on an NVidiaTitan X, i.e., 60fps.

It indicates the model you released which contains only 1 refinement level (i.e., StereoNet-8x-single) can be faster than 60fps.
However, the speed is only about 20fps when I run your code on my NVIidiaTitan X.

Is there something I have missed? Please help me figure out. Thanks

作者您好，拜读您的代码，发现只有在SceneFlow数据集上预训练的代码，没有找到对应的在KITTI上微调和测试的代码，不知道作者能不能公布一些，万分感谢啦！！！

蟹蟹~~~

How to set the network input parameter maxdisp

I used the pretrained model. The testing dataset is the default flyingthings3D TEST.
After 'outputs = model(imgL, imgR)' I got len(outputs)=2 , which caused the line 'for x in range(stages): ... output = torch.squeeze(outputs[x], 1)' crashed. How can I solve this, please? Thanks.

I am trying to train the network and I am getting the following errors

Traceback (most recent call last):
  File "./main8Xmulti.py", line 22, in <module>
    from models.StereoNet8Xmulti import StereoNet
ModuleNotFoundError: No module named 'models.StereoNet8Xmulti'

Changing from models.StereoNet8Xmulti import StereoNet to from models.StereoNet_single import StereoNet fixes that error, but then I get this error

Traceback (most recent call last):
  File "./main8Xmulti.py", line 365, in <module>
    main()
  File "./main8Xmulti.py", line 170, in main
    train(TrainImgLoader, model, optimizer, log, epoch)
  File "./main8Xmulti.py", line 223, in train
    losses[idx].update(loss[idx].item() / args.loss_weights[idx])
IndexError: list index out of range

loss only has 2 elements in the list, but losses has 4. How can I fix this?

(xstereolab) root@c792cba59ad3:/mnt/X-StereoLab-master# python3 tools/train_net_disp.py --cfg ./configs/config_disp.py --savemodel ./outputs/MODEL_NAME -btrain 4 -d 0-3 --multiprocessing-distributed
Traceback (most recent call last):
File "tools/train_net_disp.py", line 25, in
from disparity.models import *
File "/mnt/X-StereoLab-master/disparity/models/init.py", line 1, in
from .stereonet import StereoNet
File "/mnt/X-StereoLab-master/disparity/models/stereonet.py", line 11, in
from dsgn.utils.bounding_box import compute_corners, quan_to_angle, \

In the original paper, we can roughly get that when using model "8x multi", EPE is 1.1 (from table 1 ) and running time is 0.015 (from table 2, which is roughly 60fps).

In your implementation, you get 30FPS using model "8x single" while getting epe 1.38
Can you explain it more ? Or maybe the running time 0.015 is't corresponding to "8x multi"?

def warp(x, disp):
"""
warp an image/tensor (im2) back to im1, according to the optical flow
x: [B, C, H, W] (im2)
flo: [B, 2, H, W] flow
"""
B, C, H, W = x.size()
# mesh grid
xx = torch.arange(0, W).view(1, -1).repeat(H, 1)
yy = torch.arange(0, H).view(-1, 1).repeat(1, W)
xx = xx.view(1, 1, H, W).repeat(B, 1, 1, 1)
yy = yy.view(1, 1, H, W).repeat(B, 1, 1, 1)
grid = torch.cat((xx, yy), 1).float()

if x.is_cuda:
    grid = grid.cuda()
vgrid = grid
vgrid[:,:1,:,:] = vgrid[:,:1,:,:] - disp

# scale grid to [-1,1]
vgrid[:, 0, :, :] = 2.0 * vgrid[:, 0, :, :] / max(W - 1, 1) - 1.0
vgrid[:, 1, :, :] = 2.0 * vgrid[:, 1, :, :] / max(H - 1, 1) - 1.0

vgrid = vgrid.permute(0, 2, 3, 1)
output = nn.functional.grid_sample(x, vgrid)
return output

如题 谢谢

You claim that this repository include PLUMENet, but here is not.

你好，在disparity.models下面找不到loss.py,是忘记上传了吗？

还有在stereonet.py找不到dsgn这个文件夹

FileNotFoundError: [WinError 3] 系统找不到指定的路径。: '/media/lxy/sdd1/stereo_coderesource/dataset_nie/SceneFlowData/frames_cleanpass/monkaa'

how to download dataset? please

作者您好，感谢您开源实现的代码

在cost volume构建的时候有些困扰。
为什么只沿着一个维度方向进行平移相减（如代码中沿width方向平移），却没有计算height方向？

for i in range(disp):
        if i > 0:
            cost[:, :, i, :, i:] = refimg_feature[ :, :, :, i:] - targetimg_feature[:, :, :, :-i]
        else:
            cost[:, :, i, :, :] = refimg_feature - targetimg_feature

Hello, I would like to ask whether the image pyramid in the hierarchical refinement networks is your own idea or the idea of communicating with the author of the paper? Looking forward to your reply

Thanks for sharing the code
Could you share the pretrained model for Kitti2015 and Kitti2012, please?

after runing the model of StereoNet8Xmulit.py in 1080Ti， which using about 0.27s with one pairs of images. It's much slower than the result that giving from you. Could you give some advices about it? Thanks a lot!

/path/to/OUTPUT_PATH这个是啥意思

Its already been a month you said that the code will be released soon but it didn't ,so can you please share the first version of the code?

Hello, is the final output of this network a list? What do the two tensors in the list represent?

I use images after polar correction but its preference is very bad, do images need other process?

how can I train a active stereonet，I can not find the corresponding code

Hello, the link to the model seems to be broken, Can you update it?

作者您好~，我运行您的代码，出现了以下的错误：
AttributeError：module 'torch.nn.functional' has no attribute 'interpolate'
请问是缺少文件吗？或者我配置的环境以及版本不正确呢？
请您指教，非常期待您的回复！~~~

万分感谢！ 蟹蟹~~