Hello, could you please provide the scale factor of depth map? When I want to convert the depth map to 3D point clouds, I cannot find this param.
Thank you

Hello, could you provide camera focal length? This is very important to my project！
Thanks a lot!
Best wishes!

    im_fname = osp.join(self.data_root, '{}.png'.format(im))
    de_fname = osp.join(self.data_root, '{}_depth.npy'.format(im))
    de_mask_fname = osp.join(self.data_root, '{}_depth_mask.npy'.format(im))

    im = np.array(Image.open(osp.join(self.data_root, im_fname)))

on the last line the im_fname is joined twice with self.data_root

From the paper, it seems that there are some invalid depths that were shown as black pixels in the depth map. I'm wondering if there is any way to handle those, perhaps using inpainting?

Hi, thanks for providing this dataset. Besides intrinsics parameters, will you also provide the extrinsics parameters of the cameras?

Hello, nice work. I noticed that the depth map is stored in .npy format, I want to know what is the bit-depth of the raw depth images?

Hi,

Thanks for the work, in our research, we need to recover the extrinsic of the virtual parameter.

I wonder how the name corresponding to the rotation of elevation angles and azimuth.

00007_00082_outdoor_020_000

is 020 asimuth and 000 corresponding to -20 elevation ?

Thanks

Hello, what is the LICENSE of this dataset? Is it meant only for academic use? Am I allowed to use it for commercial purposes? Please add a license file to clarify this.

normal data doesn't exist for train/indoors/scene_00000/scan_00007 ...

As #3 mentioned, depth around edges are noisy. Which is a common case in depth estimation and could have been solved.

As is known, depth sensor return invalid value around edge, which is often shown as black hole, instead of random points.

However, if the incomplete depth map is resized with bi-linear interpolation, the pixels around the holes would be resampled with valid depth and zeros-values in the hole, making the new depth vary from 0 to the ground truth, which is the case here

So if is there raw data of the depth sensor? And we can try to fix the bug by reprocess the depth map.

Hi. I notice that the images have been produced with different vertical and horizontal focal lengths. The images are slightly vertically stretched, their size should be (1024 x 735) to have square pixels.

When running the baselines, did you feed the images as-is or did you correct for this?

thanks

thanks for your great work，but I have a question that how can I train with the dataset while there are many zeros in the depth map, in NYUv2 dataset, there is no zero in depthmap, and how should I deal with the pixels without depth? thanks again

Both "intrinsics.txt" and the technical report claim a horizontal FoV of 60° and a vertical FoV of 45°. However, the point clouds corresponding to such FoV appear distorted. For example, the surfaces of the corner are apparently not orthogonal. The provided FoV is larger than expected.

I have examined a few cases where surfaces that are obviously supposed to be orthogonal and then fit the optimal value of FoV to make those surfaces orthogonal. The values I obtain are 42.3° horizontally and 32.4° vertically, with a standard deviation of less than 1°. The point cloud looks visually better.

Since the depth maps and surfaces look smooth and accurate, such an error is likely introduced in a post-processing step. Moreover. the FoV of 60°x45° does not imply a consistent focal length in the X-Y axes, as tan(30°):tan(22.5°) is far from the 4:3 aspect ratio. Would the authors please check the FoV after cropping?

Here are some side-by-side comparisons between the point clouds with the original FoV (60°) on the left and the corrected FoV (42.3°) on the right. All of them are viewed from the top.

The readme mentions a test set:

We have released the train and validation splits of DIODE depth and DIODE normal, including RGB images, depth maps, depth validity masks and surface normal maps. Test set is coming soon.

Do you have an estimate for when it will be released? Thanks!

Hi, thanks for the dataset. However, when visualizing the data as point cloud, points on the intensity edges are very noisy, which suggests that the depth values are wrong at those points? Any idea how to fix this?

Thanks for a great work and dataset,
Can you please give details of normalization of the ground truth depth

Hi! Thanks for your work, it is a very interesting dataset for such an important problem!

I wonder, is it possible somehow to get the good approximation of intrinsics of the camera that this dataset was taken with?

It may be crucial for some tasks and not that hard to evaluate using the checkerboard printed pattern.

Best regards.