matsuren / omnimvs_pytorch Goto Github PK

Hi,

Thank you for sharing the code, I think they've made a mistake in their datasets. For cameras' extrinsic, the config.yaml is more corresponding than the poses.txt.
For example, pose_cam2 should be : [0.000000, 1.570796326794897, 0.000000, 60.00000, 0.000000, 0.000000] (should in centimetres btw)
Rather than : [0.000000 1.570796326794897 0.000000 20.000000 0.000000 -20.000000]

Cheers

Now the given gt depthmap is .tiff format. But I find that using the official code to read, the depthmap is 3-channels and uint8 data type. Could you offer the PNG version you previously download?

Hi,

I modified your implementation for applying binocular systems, I hope you don't mind my borrowing.

In haste

The loss value gets saturated to 0.0010 after one epoch and does not change after that. How can I resolve this issue? Please help

In the dataset, There are several images' name was not matched with omnithings_train.txt. For example, In textfile, it named 00002.png but image name is 0002.png.
Change text file or image file name to escape this problem.

Thank you very much for sharing.
I encountered a problem that the resulting predicted disparity maps were completely black, while training the data without pre-trained weight file. By printing the data, I found that all the data in tensor "gt_invd_idx" from training dataset were 0, and the data in tensor "pred" which is the output of network model was 0~1 . Can I ask what is wrong? Thank you!

Thank you very much for sharing.
I encountered a problem that the loss keep 1.0000 or 0.5000 when I train the network using these default parameters. I try to modify learn rate but it's not work.

Hello, I want to know where the dataset should be placed, why it keeps saying that I don't have a DATA_DIR file.
Thank you very much for your reply

hi, I want to test this network but i found download this dataset is to slow, can you give a pre-trained model so I could test on my data, thanks a lot :-)

Hi, I have a trouble with omnihouse groundtruth. Sorry for ask you, but I cant contact with the author of dataset.

I download the omnihouse dataset from omnistereo, but the groudtruth I got is in tiff format, as follows:

The value is arroud in 0-255, not 0-65500, as described in the dataset.

And I have tried to use your code:

def load_invdepth(filename, min_depth=55):
    '''
    min_depth in [cm]
    '''
    invd_value = cv2.imread(filename, cv2.IMREAD_ANYDEPTH)
    invdepth = (invd_value / 100.0) / (min_depth * 655) + np.finfo(np.float32).eps
    invdepth *= 100  # unit conversion from cm to m
    return invdepth

but I got wrong result.

Is there something I was wrong? If I download wrong groudtruth, could you tell me where to find the right one?

Thanks a lot!

The parameters (image size, output depth size, the number of the disparity) can't be set to the ones reported in their original paper due to GPU memory consumption. Does anyone know how to reduce GPU memory consumption for this model?

Thanks,

I have a one question about depth image format.
I cannot find depth_train_640/ in omnithings directory which is downloaded from here. But, I can see the gt depth images in Omnidirectional Stereo Dataset. This depth format looks tiff.
Your depth_train_640 meant to be the above depth images? In this case, should I convert from tiff to png by myself?
BTW, thank you for all of your awesome work!

Thanks for your contribution!
Have you tried your implementation in Synthetic Urban Datasets? I found the result is poor with your default configurations. Can you provide a training config. file or a pretrained model on this dataset?

Hi,
I would like to know under what license you would like to publish this code, to know if/how it can be used.

Thanks!