Add surfels logic

[gchhablani]

No description provided.

[maturk]

Can you git merge main into your current branch. Also perhaps running some evaluations, e.g. on the provided mushroom dataset and mushroom mesh eval script could be interesting to see the differences between dn-splatter and the proposed new losses here.

[maturk]

Tested on the "vr_room" sequence from mushroom dataset:

Metric	Surfels	DN-Splatter
Acc ↓	0.0289	0.0302
Comp ↓	0.0264	0.0258
C-L1 ↓	0.0276	0.0280
NC ↑	0.8611	0.8768
F-score ↑	0.8410	0.8332

Instructions for reproducing:

• train dn-splatter:

ns-train dn-splatter --max-num-iterations 20000 --pipeline.model.use-depth-loss True --pipeline.model.sensor-depth-lambda 0.2 --pipeline.model.use-depth-smooth-loss True --pipeline.model.use-normal-loss True --pipeline.model.normal-supervision mono mushroom --data [path_to_vr_room]

• train surfels:

ns-train dn-splatter --max-num-iterations 20000 --pipeline.model.use-depth-loss True --pipeline.model.sensor-depth-lambda 0.2 --pipeline.model.use-depth-smooth-loss True --pipeline.model.use-normal-loss False --pipeline.model.use-nd-loss True --pipeline.model.use-opacity-loss True --pipeline.model.normal-supervision mono mushroom --data [path_to_vr_room]

• extract meshes with the command:

gs-mesh dn --load-config [path to config] --output-dir [path to output]

• mushroom mesh evaluation:

python dn_splatter/eval/eval_mesh_mushroom.py --gt-mesh-path path_to/room_datasets/vr_room/ --pred-mesh-path path_to_predicted_mesh/DepthAndNormalMapsPoisson_poisson_mesh.ply