Some bad results in NeRF Synthetic dataset
ZyyZyy06 opened this issue · 4 comments
I train with all the default parameters, but it fluctuates a lot on the drums data. The results of the three training sessions differed by 2 on PSNR. After looking at the rendered results, it was found that some of the results produced severe occlusion.
How did you train the scenes? Do you mean the optimized results are random?
Just use the default command.
CUDA_VISIBLE_DEVICES=1 python train.py --eval -s /data/xxx/dataset/nerf_synthetic/drums -m /data/xxx/lab_output/origin_2dgs_output/drums2
Looks like an artifact of the background area error. A similar situation exists with other synthetic data.
We haven't thoroughly tested the regularization for nerf synthetic dataset yet. One solution is to use the render_alpha for masking the normal loss in the background regions.
Ok, thanks for the answer, it's just a small flaw in the background area on the synthetic dataset that makes me feel confused