EZ-SP: Fast and Lightweight Superpoint-Based 3D Segmentation

Louis Geist, Loic Landrieu, Damien Robert · ICRA 2026 · 2026

DOI arXiv Code

Status

To read

Topics

backbones, budgeted-inference

Feeds

P1 · ICCV 2027

What they did

Replaces the CPU-bound partition step of superpoint pipelines with a learnable, fully GPU partition that is 13 times faster; with a lightweight superpoint classifier the whole pipeline fits in under 2 MB of GPU memory and runs in real time on multi-million-point scenes.

My note

Learns efficient GPU-based superpoint partitions. Essential reading if we pursue grouping or tokenization.

Code

Repository

drprojects/superpoint_transformer

Framework

PyTorch

Licence

MIT

My status

Not tried

Note

EZ-SP is released in the Superpoint Transformer repository.

BibTeX

@article{geist2026-ez-sp,
  title = {{EZ-SP: Fast and Lightweight Superpoint-Based 3D Segmentation}},
  author = {Louis Geist and Loic Landrieu and Damien Robert},
  journal = {ICRA 2026},
  year = {2026},
  doi = {10.1109/ICRA57385.2026.11696805},
  eprint = {2512.00385},
  archivePrefix = {arXiv}
}