SoftBench / README.md
KCB1's picture
Document the complete SoftBench v1 benchmark
87a7475 verified
|
Raw History Blame Contribute Delete
3.97 kB
metadata
pretty_name: SoftBench
tags:
  - 3d
  - gaussian-splatting
  - non-rigid-deformation
  - benchmark

SoftBench

SoftBench evaluates non-rigid Gaussian splat editing from sparse 3D point pairs. Each scene provides a canonical source, ordered source and target controls, and calibrated views of the target deformation.

Split Scenes Edits Target views per edit
SoftBench-Single 8 8 90
SoftBench-Multi 4 15 75

Single adapts eight NeuralEditor scenes: Chair, Drums, Ficus, Hotdog, Lego, Materials, Mic, and Ship. Multi contains Mechanical Arm, Plush Bear, Potted Plant, and Rubber Duck. The first three have four edits each and Rubber Duck has three. Each edit provides 16, 32, 64 and 128 control pairs. Multi includes the corrected Potted Plant correspondences.

Download

index.json lists the scene archives and their expected sizes and view counts. Download and extract the archives into the same directory to obtain single/<scene>/ and multi/<scene>/.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="KCB1/SoftBench",
    repo_type="dataset",
    revision="v1",
    local_dir="SoftBench",
    allow_patterns=["index.json", "single/*.tar", "multi/*.tar"],
)

Scene files

  • source/point_cloud.ply is the canonical source checkpoint. source/tsdf_surface.ply and source/integration_grid.npz supply the source geometry and integration grid used by SoftGS.
  • source/train/, source/test/, and the Single source/val/ contain source RGB or RGBA and matching cameras.
  • controls/<budget>/source.npy and controls/<budget>/<edit>.npy provide corresponding (N, 3) points in source coordinates.
  • edits/<edit>/transforms.json pairs each target image with a Blender camera-to-world matrix in the Gaussian checkpoint's world frame. Camera angles are in radians. Single target alpha is the RGBA image's alpha channel. Multi target alpha is the separate image specified by alpha_path.
  • source/tsdf_cameras.json in Multi lists the source camera poses and dimensions used to fuse the cached TSDF. Single uses source/train/transforms.json for fusion.

Run the SoftGS code from an extracted scene directory:

softgs build --model source/point_cloud.ply --surface source/tsdf_surface.ply \
    --measure source/integration_grid.npz --source controls/128/source.npy --output operator.npz
softgs edit --model source/point_cloud.ply --operator operator.npz \
    --target controls/128/deform1.npy --output edited.ply

Use deform0.npy for Single and the available edit name for Multi. The editor consumes the source assets and sparse controls. Render the edited checkpoint at the target cameras for evaluation.

Evaluation

Single uses full-image RGB metrics on white at 800 × 800. Multi uses the native target resolution and a target alpha mask thresholded at 0.5. Multi PSNR measures RGB squared error on that mask. Multi SSIM normalizes the local 11 × 11 Gaussian moments (sigma 1.5) by foreground weight and averages over foreground centers and channels. Multi LPIPS uses VGG v0.1 with RGB in [0, 1], white outside the foreground and area-resampled foreground weights on the feature grids. Both splits compute silhouette IoU over the full image with predicted opacity and target alpha thresholded at 0.5. Average views within an edit, then weight edits equally within each split.

Credits

The Single scenes are adapted from NeuralEditor: Editing Neural Radiance Fields via Manipulating Point Clouds by JK Chen, Jipeng Lyu and Yu-Xiong Wang (CVPR 2023).

Multi uses derived benchmark assets based on Industrial robot arm by olofunot (CC BY 4.0), Teddy bears by hectopod (CC BY 4.0), Clay Duck by bradacvojtech (CC BY 4.0), and Potted Plant 03 by Prince Obrymec (Sketchfab Standard). The assets were prepared with new camera views, sparse control pairs and Blender lattice edits.