RAP

Implementation of the paper:

RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing

This repository provides the evaluation/pruning component of RAP. It removes the rendering engine, Gaussian model, and scene manager from the original 3DGS implementation and enables direct pruning from a PLY file without rendering.

The code can be used as a lightweight plug-and-play module for:

• 3DGS primitive pruning
• Compression preprocessing
• Efficient transmission pipelines

---

Requirements

• Python 3.x (recommended 3.11)
• CUDA Toolkit (recommended 12.x)
• PyTorch (must match CUDA version)

---

Installation

Create environment:

conda create -n rap python=3.11
conda activate rap

Install PyTorch (example for CUDA 12.x):

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

Optional dependencies (recommended)

If you want GPU KNN acceleration (cuVS):

pip install cupy-cuda12x
pip install cuvs-cu12 --extra-index-url=https://pypi.nvidia.com

If cuVS cannot be installed, use CPU-only dependencies:

pip install plyfile scipy

---

Usage

Arguments

• --ply_path (required)
  Path to input Gaussian PLY file.

• --output_ply_path (required)
  Path to output pruned PLY file.

• --keep_percent (required)
  Percentage of primitives to retain (0¨C1).
  Example: 0.8 keeps 80% Gaussians.

• --input_dim (optional, default=15)
  Input feature dimension.

• --knn_k (optional, default=128)
  Number of neighbors for local statistics.

• --knn_method (optional)
  KNN backend:

  • ivf ¡ª fastest GPU approximate search
  • brute_force ¡ª exact GPU search
  • ckdtree ¡ª CPU exact search (recommended if CUDA/cuVS unavailable)

• --net_weights_path (optional)
  Path to pretrained network weights.
  Default: net_weights/net_f15.pth.

• --sh_degree (optional, default=3)
  Spherical harmonics degree.

• --data_device (optional, default=cuda:0)
  CUDA device index.

---

Example

python prune_percent.py \
    --ply_path /path/to/input.ply \
    --output_ply_path /path/to/output.ply \
    --keep_percent 0.8

---

Notes

• This repo focuses on importance prediction and pruning only.
• Rendering-based evaluation is intentionally removed.
• Designed for fast post-hoc 3DGS compression and analysis.

---

Citation

If you find this project useful, please cite:

@inproceedings{yang2026rap,
  title={RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing},
  author={Yang, Kaifa and Yang, Qi and Xu, Yiling and Li, Zhu},
  booktitle={CVPR},
  year={2026}
}

---