Bad result on my own dataset

[Jason-YangJX]

Thanks for your great work! Training on a synthetic image dataset generated by an exsiting mesh with the same cam settings in "Instant-dataset/transforms.json", I can't get a complete generating mesh. So I think there must be something wrong.
According to the ray sampling strategy in "nerf.utils.get_rays" , only few pixels belonging to the human body are used. Instead, I generate a body bounding box based on mask and sample within the bounding box and increase hyperparameters “--num_rays”. Is there any problem with doing this?
Otherwise, could you please explain how "nerf.provider.nerf_matrix_to_ngp" works?
Finally, I'm also expecting to get the tips mentioned in #3 (comment)