Problem of produced point cloud

[Riser6]

Hi, Lihe，congratulations on this great work and thanks for releasing the code for the community. Then in the process of using this project, I found that the point cloud it produced was not that good(I used the image from the ScanNet dataset as test example) with the code in https://github.com/LiheYoung/Depth-Anything/blob/main/metric_depth/depth_to_pointcloud.py, but I found the website(https://huggingface.co/spaces/Xenova/depth-anything-web) produced more reasonable pointcloud, but still not precise enough. Could you possibly know what might have caused this?

Produced by python code:

Produced by the website:

[mikami520]

@Riser6 Hi, I have the same question here, did you solve the problem? I use the same code to convert depth map to point cloud but with two different depth maps (one is from ZoeDepth, another is from DepthAnything; In fact, both gives the disparity instead of depth), the results from ZoeDepth made more sense compared to DepthAnything, do you have any suggestions?