Reference Twice: A Simple and Unified Baseline for Few-Shot Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Few-Shot Instance Segmentation (FSIS) aims to detect and segment novel object classes using minimal labeled examples.
• Existing Region Proposal Network (RPN)-based methods suffer from overfitting and complex spatial correlation strategies.
• Dual-branch models in FSIS often lose spatial information during prototype generation.

Purpose of the Study:

• To introduce a unified framework, Reference Twice (RefT), for Few-Shot Instance Segmentation (FSIS).
• To address overfitting and spatial information loss issues in existing FSIS methods.
• To improve the performance and extensibility of FSIS models, particularly in incremental settings.

Main Methods:

• Developed a novel transformer-based baseline to mitigate overfitting in FSIS.
• Implemented a dual enhancement strategy for query features at both feature and query levels using cross-attention.
• Introduced a class-enhanced base knowledge distillation loss for incremental FSIS extension.

Main Results:

• The Reference Twice (RefT) framework demonstrates improved performance on the COCO dataset across various few-shot settings.
• Achieved significant performance gains over state-of-the-art methods, including +8.2% with 10-shots and +9.4% with 30-shots.
• The proposed transformer-based approach effectively avoids overfitting and simplifies spatial correlation mechanisms.

Conclusions:

• The Reference Twice (RefT) framework offers a promising new direction for Few-Shot Instance Segmentation.
• The method successfully enhances feature representation and query understanding without complex spatial interactions.
• RefT provides a flexible and effective solution for both standard and incremental FSIS tasks.