The blurring of the edges in the depth map leads to noisy points when merging the 3D point cloud.

[yjb6]

I greatly appreciate this wonderful work and the shared dataset.

When I tried to obtain a point cloud from the depth map, I found that there were many noisy points. For example, from the 0000_nd.exr in render_data_examples:

I believe this is caused by erroneous information at the edges of the depth map, and noticeable aliasing can be found at the edges of the obtained depth map, as mentioned in #32

Is there a good solution to this problem?

[yjb6]

blurring, not aliasing

[HuangZiYang2000]

Hi, Is there any solution by now?

[TianjiJiang]

I believe the blurring is most likely caused by interpolation during image downsampling. This issue can often be mitigated by using a synchronized downsampling mask to identify reliable pixels. Unfortunately, the dataset does not provide such a mask directly.

However, I found that the provided .png images are 4-channel RGBA images. The alpha channel can, to some extent, be treated as a downsampling mask.

Specifically, you can read the full image and extract the alpha channel to determine which pixels are valid. For example:

import cv2

img = cv2.imread('image.png', cv2.IMREAD_UNCHANGED) # BGRA image with 4 channel
alpha = img[:, :, -1]
valid_mask = alpha > 254 

However, this approach also has its limitations. If there are semi-transparent objects in the dataset, they might be unintentionally masked out. Therefore, this method is not perfect. It would be ideal if the dataset authors could release either the full-resolution images or the corresponding masks, which would fundamentally resolve the issue.