핵심 개념
IVLOD introduces ZiRa to adapt VLODMs incrementally while preserving zero-shot generalization.
초록
Abstract: IVLOD introduces ZiRa to adapt VLODMs incrementally while preserving zero-shot generalization.
Introduction: VLODMs excel in zero-shot recognition but struggle in specialized domains, necessitating incremental learning.
Challenges: Catastrophic forgetting and maintaining zero-shot generalizability are key challenges in IVLOD.
Methodology: ZiRa introduces RDB and ZiL to address challenges in IVLOD effectively.
Experiments: ZiRa outperforms existing methods in preserving zero-shot generalization and adapting to new tasks.
Conclusion: ZiRa offers a memory-efficient solution for IVLOD with superior performance.
통계
Comprehensive experiments on COCO and ODinW-13 datasets.
ZiRa outperforms CL-DETR and iDETR by 13.91 and 8.71 AP, respectively.
인용구
"ZiRa effectively safeguards the zero-shot generalization ability of VLODMs while continuously adapting to new tasks."
"ZiRa eliminates the need for saving the entire model copy for distillation or maintaining exemplars for replaying."