Skip to content

run&eval.md

Latest commit

 

History

History
62 lines (55 loc) · 2.69 KB

run&eval.md

File metadata and controls

62 lines (55 loc) · 2.69 KB

Run & Evaluate

The detailed setting of each step on GauU-Scene and MatrixCity Dataset can be found in ./scripts/run_citygs_SCENE.sh.

  • If applying 3DGS, please change model and renderer as done in configs/citygs_lfls_sh2_trim.yaml over configs/citygsv2_lfls_sh2_trim.yaml.
  • To obtain full model, adjust SH degree to 3 and enable diable_trimming in renderer.
  • You can also run on Mill19 or UrbanScene3D dataset by setting configs according to that of branch V1-Original.

A. Train coarse model

# $COARSE_NAME is the name of coarse model training config file.
python main.py fit \
        --config configs/$COARSE_NAME.yaml \
        -n $COARSE_NAME \

Tips: If you want to accelerate this part, you can use 7K iteration or lower resolution (Remember to adjust the resolution of estimated depth accordingly). But the overall performance may drop, as validated in our CityGaussianV2 paper and issue Linketic#124.

B. Model Partition and Data Assignment

# $NAME is the name of parallel tuning config file.
# to partition in contracted space, add `--contract`
# to partition after xy plane is aligned to ground plane, add `--reorient`
python utils/partition_citygs.py --config_path configs/$NAME.yaml --force

The command also visualizes block division and data partitioning, as like below: partition

C. Finetune model parallelly and merge

python utils/train_citygs_partitions.py -n $NAME
python utils/merge_citygs_ckpts.py outputs/$NAME

D. Evaluate Rendering Performance

# For finetuned model, since the split mode and eval ratios are changed for per-block tuning, the parameters have to be reappointed. Please see the script for details.
python main.py test \
        --config outputs/$NAME/config.yaml \
        --save_val \
        --test_speed \  # calculate FPS

E. Mesh extraction and evaluation

python utils/gs2d_mesh_extraction.py \
        outputs/$NAME \
        --voxel_size $VOXEL_SIZE \
        --sdf_trunc $SDF_TRUNC \
        --depth_trunc $DEPTH_TRUNC \

python tools/eval_tnt/run.py \
        --scene your_gt_pcd \
        --dataset-dir data/geometry_gt/your_scene \
        --transform-path data/geometry_gt/your_scene/transform.txt \
        --ply-path "outputs/$NAME/fuse_post.ply"

F. Compression

python tools/vectree_lightning.py \
        --model_path outputs/$NAME \
        --save_path outputs/$NAME/vectree \

If you use SH degree of 2, please add --sh_degree 2. And if you want to decode the quantized result to checkpoint, add --no_save_ply. If 3DGS is used, add --gs_dim 3.