HiT: Hierarchical Transformers for Unsupervised 3D Shape Abstraction

Aditya Vora, Lily Goli, Andrea Tagliasacchi, Hao Zhang

This repository contains the official implementation of HiT: Hierarchical Transformers for Unsupervised 3D Shape Abstraction.

Installation Steps

1. Clone the Repository:

   git clone --recursive https://github.com/aditya-vora/HiT
   cd HiT

2. Set Up the Conda Environment:

   conda create -n hit python=3.10;
   conda activate hit

3. Install the dependencies inside the conda environment:

   pip install -r requirements.txt

Training

Train HiT on ShapeNet-formatted data (see preprocess/ for how to build the required .hdf5 files):

sh scripts/train_all_cats.sh

or invoke train.py directly, e.g.:

python train.py --train --data-dir <path-to-data> --dataset-name shapenet --cats all --exp-name my_experiment

Evaluation

eval_seg.py loads a trained checkpoint (--ckpt) and evaluates it on the test split in one of five modes, selected by passing exactly one of the following flags:

Flag	Output
--recon	Colored part + full-shape meshes for a single shape (--shape-id), via marching cubes.
--mesh	Colored part + full-shape meshes for every shape in the test set.
--pointcloud	A colored, per-part-segmented point cloud (.ply) for every test shape.
--iou	Per-shape, per-category, and overall part-segmentation mIoU against ground-truth labels.
--cd	Per-category and overall Chamfer Distance / F-score against the ground-truth surface.

sh scripts/eval_all_cats.sh

or directly, e.g. to compute segmentation mIoU:

python eval_seg.py --iou --ckpt <path-to-checkpoint.pt> --data-dir <path-to-data> --dataset-name shapenet --cats all --splits test

Results are written under <exp-dir>/<exp-name>/<NNN>/.

TODO

• Release training code.
• Release data processing scripts.
• Release evaluation scripts.
• Testing.

BibTeX

      @inproceedings{vora2026hit,
        title={HiT: Hierarchical Transformers for Unsupervised 3D Shape Abstraction},
        author={Vora, Aditya and Goli, Lily and Tagliasacchi, Andrea and Zhang, Hao},
        booktitle={2026 International Conference on 3D Vision (3DV)},
        pages={1598--1607},
        year={2026},
        organization={IEEE}
      }