Informativeness-Aware Layer Freezing and Sample Replay for Efficient Online Continual Object Detection 


Vol. 15,  No. 4, pp. 324-333, Apr.  2026
https://doi.org/10.3745/TKIPS.2026.15.4.324


PDF
  Abstract

Modern object detectors deployed on drones and robots must continually adapt to new classes and changing domains under tight computational and memory budgets. Achieving continual adaptation efficiently requires learning only from informative signals in parameters and in data. Yet existing continual detection approaches either update all layers uniformly, wasting compute on saturated representations, or rely on costly sample-selection heuristics that demand extra inference. Thus, we present an informativeness-aware continual object detection framework that unifies layer freezing and frequency-loss guided sample replay into a single cost-efficient objective. At each step, we selectively updates only the layers predicted to yield the highest information per computation while probabilistic sampler prioritizes hard-but-under-trained samples by combining sample usage frequency and exponential-moving-average updated detection loss. This joint design ensures that both computation and memory are allocated to the most informative components of learning, enabling detectors to adapt rapidly without unnecessary cost. Our method consistently outperforms existing baselines on VOC, MS-COCO (Class-Incremental Learning) and SHIFT (Domain-Incremental Learning) benchmarks using YOLO- and DETR-style detectors. The proposed strategy achieves up to 39% reductions in training FLOPs with improved average precision and enhanced stability across class and domain shifts, establishing informativeness as a unifying principle for computationally and memory-efficient online continual object detection.

  Statistics
  Cite this article

[IEEE Style]

T. Kim, M. Lee, J. Park, C. Lee, C. Lee, M. Kim, D. Lee, J. Choi, "Informativeness-Aware Layer Freezing and Sample Replay for Efficient Online Continual Object Detection," The Transactions of the Korea Information Processing Society, vol. 15, no. 4, pp. 324-333, 2026. DOI: https://doi.org/10.3745/TKIPS.2026.15.4.324.

[ACM Style]

Taeheon Kim, Minjae Lee, Jonghyun Park, Chaeeun Lee, Changdae Lee, Mingi Kim, Dongseok Lee, and Jonghyun Choi. 2026. Informativeness-Aware Layer Freezing and Sample Replay for Efficient Online Continual Object Detection. The Transactions of the Korea Information Processing Society, 15, 4, (2026), 324-333. DOI: https://doi.org/10.3745/TKIPS.2026.15.4.324.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals