Open zhj1013 opened 7 months ago
Getting similar issues where the reconstructed object is VERY smoothed.
Here are the params I ran with:
Below is are several sample images
CONFIG=<used exact output of the COLMAP step>
EXPERIMENT=my_run
GROUP=my_group
NAME=my_name
GPUS=1
torchrun --nproc_per_node=${GPUS} train.py \
--logdir=logs/${GROUP}/${NAME} \
--config=CONFIG \
--show_pbar
Here are my WandB results
A. In the research paper I saw that bumping hash encoding level from 4 to 16 significantly improves details. Is this what could be causing this overly smooth output?
If so, at what step (COLMAP, Nueralangelo, Isosurface extraction) should this level be set and where in the config?
Is it the model.object.sdf.encoding.coarse2fine.init_active_level
param?
B. Could it be related to the training never reaching the sufficient activation level? The default config set max_iter at 500k, so I wonder if this has limited the models ability to extract more exact surfaces?
I attempted 3D reconstruction using custom data but failed. My dataset was extracted from a video of an object on a turntable shot with a smartphone, using colmap. The results from visualize_colmap.ipynb and visualize_transforms.ipynb appear to be normal. The wandb val/vis/opacity seems off. Can someone help me figure out what the issue might be?