How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?

Tuan Anh Tran, Duy M. H. Nguyen, Hoai-Chau Tran, Michael Barz, Khoa D. Doan, Roger Wattenhofer, Ngo Anh Vien, Mathias Niepert, Daniel Sonntag, Paul Swoboda · NeurIPS 2025 · 2025

arXiv Project page

Status

To read

Topics

backbones, budgeted-inference

Feeds

P1 · ICCV 2027

What they did

Shows that tokens in 3D point cloud transformers are highly redundant and introduces gitmerge3D, a globally informed graph token merging method that removes up to 90–95% of tokens while keeping competitive accuracy.

My note

Adaptive token merging for point-cloud transformers. A central compression comparator.

BibTeX

@inproceedings{tran2025-gitmerge3d,
  title = {{How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?}},
  author = {Tuan Anh Tran and Duy M. H. Nguyen and Hoai-Chau Tran and Michael Barz and Khoa D. Doan and Roger Wattenhofer and Ngo Anh Vien and Mathias Niepert and Daniel Sonntag and Paul Swoboda},
  booktitle = {NeurIPS 2025},
  year = {2025},
  eprint = {2511.05449},
  archivePrefix = {arXiv}
}