search

BaowenZ / RaDe-GS

RaDe-GS: Rasterizing Depth in Gaussian Splatting
Other
328 stars 7 forks source link

Poor results on custom dataset when distortion loss and normal loss starts #21

Open mbernier-arcturus opened 1 week ago

mbernier-arcturus commented 1 week ago

Hi! Similar to #15 I get very poor results on my own dataset as soon as the depth distortion loss and normal loss kicks in (iteration 15k), it just gets extremely blurry. I tried adjusting the lambda distortion from 100 to 1, and it's similar.

Do you have any suggestions on what could possibly cause this?

I am using -r 2 --use_decoupled_appearance and my parameters are:

self.iterations = 30_000 self.position_lr_init = 0.00016 self.position_lr_final = 0.0000016 self.position_lr_delay_mult = 0.01 self.position_lr_max_steps = 30_000 self.feature_lr = 0.0025 self.opacity_lr = 0.05 self.scaling_lr = 0.005 self.rotation_lr = 0.001 self.appearance_embeddings_lr = 0.001 self.appearance_network_lr = 0.001 self.percent_dense = 0.01 self.lambda_dssim = 0.2 self.lambda_distortion = 100 self.lambda_depth_normal = 0.05 self.densification_interval = 100 self.opacity_reset_interval = 3000 self.densify_from_iter = 500 self.densify_until_iter = 15_000 self.regularization_from_iter = 15_000 self.densify_grad_threshold = 0.0002

BaowenZ commented 1 week ago

Hi! Are the results on DTU dataset still poor? If the results are poor, you could try https://github.com/BaowenZ/RaDe-GS/issues/4#issuecomment-2185893564

mbernier-arcturus commented 6 days ago

Thanks for the reply! The results for the DTU dataset are (very) good... Mine is a very different scene, more large scale, with obstruction. It work well (not perfect) with 2DGS and 3DGS, so that's why I'm wondering what would make it disintegrate here (I did try to reinstall/setup the submodule, but it was a fresh install)