city-super / Octree-GS

Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians

Other

518 stars 36 forks source link

Not able to replicate results #35

Closed voidcf closed 2 months ago

voidcf commented 2 months ago

Hi, thanks for the great work! Lately, I've been trying to replicate the results reported by Octree-GS in their paper for the Block_small scene, but the maximum PSNR I've gotten so far was ~18.

I used the block_small scene by taking the files only from the aerial folder and the transforms.json inside it as mentioned here and here
I also made changes to the cuda rasterizer as suggested here and recompiled it
The GPU that I am using is a Quadro RTX 6000 (24G)

Am I missing something here, please do let me know... I've attached the log file for your reference as well.

jianglh-WHU commented 2 months ago

can you show the poses of training images and the init point cloud?

voidcf commented 2 months ago

I've used the same point cloud that you mentioned here and poses are from the transforms.json file in MatrixCity/aerial_street_fusion/aerial/transforms.json

jianglh-WHU commented 2 months ago

can you visualize it?

voidcf commented 2 months ago

16, my point cloud is looking similar to this.

jianglh-WHU commented 2 months ago

In MatrixCity, we conducted an experiment that involved fusing data from both aerial and street views. To represent this data, we provided aerial_street_fusion, which is discussed in more detail in issue #30. However, for the Octree-GS experiment, we utilized aerial images from the block_small region. For your convenience, we have made available the point cloud and pose for your use.

voidcf commented 2 months ago

Yeah it works fine... some bug at my end. Thanks a lot👍