Add Dsine Surface Normal Option

[pablovela5620]

Add the option to use the DSINE surface normal estimator

To test out, just run

pixi run example-polycam

This will

1. Download an example polycam scan
2. Convert said polycam scan -> nerfstudio format
3. Generate surface normals based on DSINE
4. Create a pointcloud similar to how its done for mushroom
5. Train using dnsplatter

• Included a script to compare different surface normal estimations. Can be run by

python dn_splatter/scripts/compare_normals.py \
   $PATH TO FIRST SURFACE NORMAL DIR \
   $PATH TO SCOND SURFACE NORMAL DIR

Heres an example of a custom dataset I collected

• Estimate surface normals from pretrain using

python dn_splatter/scripts/normals_from_pretrain.py --model-type dsine

DSINE predicts normals in a range of -1 to 1 in camera coordinate system LUB (+X left | +Y up | +Z Back), I convert them to RUB.
I left the current convention names as they are but they appear to be backwards (no issues with functionality, just naming convention errors #9

• Also added a small change to mushroom pointclouds utils to allow generation of pointclouds for polycam datasets generated by nerfstudio, this allows for using datasets generated via polycam to not have to use random gaussian initialization which leads to better convergence

python dn_splatter/data/mushroom_utils/pointcloud_utils.py

• added normal visualization and metrics to get_image_metrics_and_images

[gaetan-landreau]

Hi ! 👋🏼
Regarding the DSINE normal orientation convention, this issue seems to admit a RDF convention (it's equivalent, RDF = -LUB). But then the conversion to RUB through diag(-1,1,1) seems a bit wrong to me then. Am I missing something ?