Official Implementation of "Spectrum AUC Difference (SAUCD): Human-aligned 3D Shape Evaluation" (CVPR 2024).

Prepartion

Install environment

sh env.sh

Download trained weights and required files from here. Unzip it and put it under assets folder.

Prepare dataset

1. Download the Shape Grading dataset from here.
2. Unzip it and put it under dataset folder.
3. Preprocess the laplacian operater and eignn decomposition of the meshes in the dataset.

   python preprocess/compute_eig.py
   

   This could take a few hours and take up to ~120GB hard drive space to store the results.

Testing SAUCD and Weighted SAUCD

python experiments/sota.py 

Training weights for Weighted SAUCD

cd train_weights
python main/train.py --config_file configs/debug.yaml