Bad result on custom dataset

[hongju91]

Hi, I am trying to run the model with my own dataset, which is similar to the NHR dataset with 80 2k cameras surrounding a person. I found the metrics stop to grow after only 3 or 4 epochs, and in the results, it seems the size of every point is big, it's clear to see each point and the space between them. Could you give some suggestion on what might be causing the problem? A bad calibration or wrong config parameters? Thanks a lot for your help.

[dendenxu]

Hi, looks like the number of points is too small. We typically use 26w (260000) points per frame for such datasets.
You can try lowering the voxel size during visual hull reconstruction to get a more dense initial point cloud estimation.

[dendenxu]

The process for extracting vhulls is detailed here.

[hongju91]

Thanks I will try that to see if I can get some better results. Besides, is there any suggestions on how to capture the images, for instance: the light condition, the percentage of image that occupied the human, etc...? Thank you!

[hongju91]

Thanks I will try that to see if I can get some better results. Besides, is there any suggestions on how to capture the images, for instance: the light condition, the percentage of image that occupied the human, etc...? Thank you!

[dendenxu]

Typically, what matters more on the dataset side is the quality of camera synchronization and camera parameters.
The lighting condition and ratio of human pixels depend more on the image quality produced by your camera sensor. As long as they're shared across these multiple views, the lighting condition shouldn't impact the reconstruction too much. As for the ratio of human pixels, it should be as large as possible if you're only interested in rendering the foreground humans & objects.