Core Concepts
Developing a novel method, CD-ViTO, enhances open-set detectors for accurate cross-domain few-shot object detection.
Abstract
The content discusses the challenges of cross-domain few-shot object detection (CD-FSOD) and introduces the CD-ViTO method to address them. It includes an in-depth analysis of datasets, benchmark creation, evaluation of various methods, and the effectiveness of proposed modules like learnable instance features, instance reweighting, and domain prompter. Results show significant improvements over existing models.
Introduction
Challenges in Cross-Domain Few-Shot Learning.
Introduction to CD-FSOD and DE-ViT model.
Methodology
Overview of CD-ViTO.
Detailed explanation of learnable instance features (MLIF), instance reweighting module (MIR), and domain prompter (MDP).
Experiments
Evaluation on different datasets using various methods.
Analysis of results for 1/5/10 shot scenarios.
Analysis
Impact of style, ICV, and IB on domain gap.
Ablation study on proposed modules: MLIF, MIR, MDP.
Conclusion
Summary of contributions to CD-FSOD field.
Stats
"CD-ViTO surpasses Meta-RCNN by 332.1% under the 10-shot setting on ArTaxOr."
"DE-ViT achieves 9.2 mAP under the 10-shot setting."
"ViTDeT-FT shows strong performance on ArTaxOr but less effective on DeepFish."
Quotes
"CD-ViTO significantly improves upon the base DE-ViT."
"Finetuning is crucial in CD-FSOD."