search

buaavrcg / LEGaussians

Pytorch Code for "LEGaussians: Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding"
https://buaavrcg.github.io/LEGaussians/
MIT License
100 stars 14 forks source link

Nan Values for mIoU and mAP #13

Open davidpengiupui opened 1 month ago

davidpengiupui commented 1 month ago

Thank you so much for your great work! I met some problems when trying to run on the "garden" scene of MipNeRF360 and really need your help. I could get normal loss after training, but when evaluating, the results are almost all Nan.

This is after training normal_loss_LEGaussian And this is after evaluation bad_result_LEGaussian

davidpengiupui commented 1 month ago

Here are the commands that I used

cd preprocess
python quantize_features.py --config configs/mipnerf360/garden.cfg

cd ..
python train.py --config configs/mipnerf360/garden.cfg

python render_mask.py --config configs/mipnerf360-rendering/garden.cfg

cp -r data/mipnerf360/garden/segmentations output/mipnerf360/garden/0/open_new_eval_softmax_s10.0_a05/ 
python eval.py --path output/mipnerf360/garden/0/open_new_eval_softmax_s10.0_a05
Chuan-10 commented 1 month ago

Hello, how about the generated masks in output/mipnerf360/garden/0/open_new_eval_softmax_s10.0_a05/? Do they appear normal?

davidpengiupui commented 1 month ago

The generated masks look not very good. For example, the wood pot: wood pot and the windows windows

Chuan-10 commented 1 month ago

You can use pdb to trace the source of the invalid value, such as the possibility of gt not being read properly or the count having a denominator of 0. Bad masks would typically lead to a low value rather than NaN.