Test-Time Few-Shot Object Detection via Dynamic Prototype Fusion.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Test-time few-shot object detection (FSOD) aims to identify novel object categories with limited examples without model fine-tuning.
• Existing FSOD methods face challenges with domain/category shift and limited data availability.
• Previous research in test-time FSOD has shown promise but requires further improvements.

Purpose of the Study:

• To propose a novel dynamic prototype fusion network (PFN) to address limitations in test-time FSOD.
• To mitigate distribution shift and limited data issues in few-shot object detection.
• To enhance the accuracy and robustness of FSOD models in real-world scenarios.

Main Methods:

• Introduced a dynamic prototype refinement method for adaptive prototype updates from support images.
• Developed a dual-level multi-scale information integration approach to fuse information across network layers and image scales.
• Utilized a mask-based preprocessing technique with segmentation labels to reduce background noise impact.

Main Results:

• The proposed PFN method demonstrated superior performance compared to state-of-the-art FSOD methods on multiple benchmarks.
• Adaptive prototype refinement effectively handled distribution shifts.
• Multi-scale information integration and mask-based preprocessing improved model discriminating capabilities and accuracy.

Conclusions:

• The dynamic prototype fusion network (PFN) offers a significant advancement in test-time few-shot object detection.
• The method effectively overcomes challenges of distribution shift and limited data availability.
• PFN shows remarkable potential for practical applications requiring rapid adaptation to new object categories.