omnihouse gt about omnimvs_pytorch

[imuncle]

I think the groundtruth can't be used now, it's broken. I also tried the method in #5 ,but I also can't train to abtain a good result. However the dataset in http://vclab.kaist.ac.kr/cvpr2021p1/ is able to use, so the problem is not in my code.

Do you still have the ground truth you downloaded in png format? Could you share it to me?

[wangyushen]

I think the groundtruth can't be used now, it's broken. I also tried the method in #5 ,but I also can't train to abtain a good result. However the dataset in http://vclab.kaist.ac.kr/cvpr2021p1/ is able to use, so the problem is not in my code.

Do you still have the ground truth you downloaded in png format? Could you share it to me?

I have the same problem as you. In the omnimvs (CVPR) has code to extract the TIFF format Ground-Truth, but the disparity map looks wrong after using. In the second layer of the current dataset, you can see the GT as the example.

[imuncle]

I think the groundtruth can't be used now, it's broken. I also tried the method in #5 ,but I also can't train to abtain a good result. However the dataset in http://vclab.kaist.ac.kr/cvpr2021p1/ is able to use, so the problem is not in my code.
Do you still have the ground truth you downloaded in png format? Could you share it to me?

I have the same problem as you. In the omnimvs (CVPR) has code to extract the TIFF format Ground-Truth, but the disparity map looks wrong after using. In the second layer of the current dataset, you can see the GT as the example.

Sorry, what do you mean by 'the second layer' ?

[wangyushen]

I think the groundtruth can't be used now, it's broken. I also tried the method in #5 ,but I also can't train to abtain a good result. However the dataset in http://vclab.kaist.ac.kr/cvpr2021p1/ is able to use, so the problem is not in my code.
Do you still have the ground truth you downloaded in png format? Could you share it to me?

I have the same problem as you. In the omnimvs (CVPR) has code to extract the TIFF format Ground-Truth, but the disparity map looks wrong after using. In the second layer of the current dataset, you can see the GT as the example.

Sorry, what do you mean by 'the second layer' ?

A TIFF format file contains multiple pictures, and the TIFF format file in this dataset contains two pictures, the first is a 3-channel picture and the second is a single channel picture. The second image is very similar to the disparity map, but it is treated as a depth map in OmniMVS (CVPR). The processed image is not like the correct disparity map.

[imuncle]

@wangyushen Great! You have solved my problem. Thanks bro!

[wangyushen]

@wangyushen Great! You have solved my problem. Thanks bro!

How did you solve this problem? Could you please share the code of data loading?

[imuncle]

The second layer is exactly the invdepth, and it does not need any more transformation.

from PIL import Image
import numpy as np
gt = Image.open('00001.tiff')
gt.seek(1)    # get the second layer
gt = np.array(gt) # this is the invdepth
# by using gt = 1/gt you can get the depth in meters

What's more, the distance between camera and the center of rig is 0.6m, not 0.2m. So the config.yaml is true, and the poses.txt is wrong.

I verify it by wraping the groudtruth from Equidistant projection to the fisheye projection. When I suppose the distance is 0.6m, I got the following result:

As you can see, the depth overlaps perfectly with the image

[matsuren]

Sorry for late reply, and thank you for great discussion!
It's better to mention in #8 about wrong extrinsic parameters. I should've been more skeptical about the extrinsic parameters.