Token Cropr: Faster ViTs for Quite a Few Tasks

Benjamin Bergner, Christoph Lippert, Aravindh Mahendran · CVPR 2025 · 2025

DOI arXiv Code

Status

To read

Topics

budgeted-inference, other

Feeds

P2 · NeurIPS 2027

What they did

Prunes Vision Transformer tokens by task relevance, learned end to end with auxiliary prediction heads that are removed after training, raising throughput across classification and dense prediction tasks.

My note

Learns task-relevant token selection, including for semantic and instance segmentation.

Code

Repository

benbergner/cropr

Framework

PyTorch

Licence

MIT

My status

Not tried

BibTeX

@inproceedings{bergner2025-token-cropr,
  title = {{Token Cropr: Faster ViTs for Quite a Few Tasks}},
  author = {Benjamin Bergner and Christoph Lippert and Aravindh Mahendran},
  booktitle = {CVPR 2025},
  year = {2025},
  doi = {10.1109/CVPR52734.2025.00910},
  eprint = {2412.00965},
  archivePrefix = {arXiv}
}