EUVIMEDEUVIMED entdecken
Crossref

Learning Diversified Primitive Prompts for Compositional Zero-Shot Learning

Ziyi Chen, Xinru Zhao, Congyan Lang, Lili Wei, Tao Wang, Yidong Li

IEEE Transactions on Circuits and Systems for Video Technology · 2025-10 · 35 (10) · 10423-10436

DOI: 10.1109/tcsvt.2025.3563606 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2025-10. Verfasst von Ziyi Chen und 5 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 65 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Beijing Nova Program, Beijing Natural Science Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Compositional Zero-Shot Learning (CZSL) aims to combine known attributes and objects as primitives for recognizing previously unseen attribute-object pairs. The essence of CZSL lies in separately modeling attributes and objects for transfer to unseen compositions. Since primitives exhibit weak entanglement in language features, Vision-Language Models (VLMs) based CZSL methods focus on prompt learning to assist primitive disentanglement in visual spaces. However, these approaches overlook the polysemy of primitives across different compositional contexts, leading the models to overly rely on seen compositions, thereby limiting their generalization capability. To tackle this issue, we propose a novel primitive disambiguation model named DPR, aiming to enhance CZSL performance by acquiring diversified primitive representations through joint-prompt learning. Specifically, DPR introduces Knowledge-Aware Hard Prompting (KAHP), which generates diverse descriptive sentences as supplementary prompts for each composition from the rich knowledge of Large Language Models (LLMs). Simultaneously, we design Sample-Adaptive Soft Prompting (SASP), which employs a lightweight neural network to generate input-guided tokens for each image. SASP ensures enhanced generalization capability in DPR’s primitive representations. We validate the effectiveness of our model through extensive experiments and achieve state-of-the-art performance on three CZSL benchmark datasets in both closed- and open-world settings.

Förderung

  • National Natural Science Foundation of China
  • Beijing Nova Program
  • Beijing Natural Science Foundation

Zitieren

APA

Chen, Z., Zhao, X., Lang, C., Wei, L., Wang, T., & Li, Y. (2025). Learning Diversified Primitive Prompts for Compositional Zero-Shot Learning. IEEE Transactions on Circuits and Systems for Video Technology, 35(10), 10423-10436. https://doi.org/10.1109/tcsvt.2025.3563606

Vancouver

Chen Z, Zhao X, Lang C, Wei L, Wang T, Li Y. Learning Diversified Primitive Prompts for Compositional Zero-Shot Learning. IEEE Transactions on Circuits and Systems for Video Technology. 2025;35(10):10423-10436. doi:10.1109/tcsvt.2025.3563606

BibTeX
@article{Chen2025,
  author = {Chen, Ziyi and Zhao, Xinru and Lang, Congyan and Wei, Lili and Wang, Tao and Li, Yidong},
  title = {Learning Diversified Primitive Prompts for Compositional Zero-Shot Learning},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2025},
  volume = {35},
  number = {10},
  pages = {10423--10436},
  doi = {10.1109/tcsvt.2025.3563606}
}
Kostenlos starten →