A robot that learns to recognize a new object should not have to re-study every object it has seen before — and should not need to keep those images around either. RECALL performs object classification in a continual/incremental fashion without rehearsal, reaching state-of-the-art results.

A frozen feature extractor is paired with a lightweight regression head trained on recall labels, so previously learned categories are preserved without storing or replaying their samples. On CORe50, iCIFAR-100 and HOWS-CL-25, RECALL outperforms other rehearsal-free continual learning approaches, which suffer from catastrophic forgetting as sequences grow.

Evaluated on HOWS-CL-25, the synthetic household object dataset built for exactly this purpose.

Publication

RECALL: Rehearsal-free Continual Learning for Object Classification

Knauer, M., Denninger, M., Triebel, R.

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022

DOI arXiv DLR elib

BibTeX

@inproceedings{knauer2022recall,
  title     = {RECALL: Rehearsal-free Continual Learning for Object Classification},
  author    = {Knauer, M. and Denninger, M. and Triebel, R.},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2022},
  doi       = {10.1109/IROS47612.2022.9981968},
  url       = {https://arxiv.org/abs/2209.14774}
}

Publication

RECALL: Rehearsal-free Continual Learning for Object Classification

Links