Transformer-Based Approach Via Contrastive Learning for Zero-Shot Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot detection (ZSD) aims to identify objects from classes not seen during training, relying on semantic information.
• Existing two-stage ZSD methods struggle with poor region proposals, inadequate consideration of inter-class correlations, and domain bias towards seen classes.

Purpose of the Study:

• To introduce Trans-ZSD, a novel transformer-based framework for improved zero-shot detection.
• To address limitations of existing ZSD methods, including poor region proposals and domain bias.

Main Methods:

• Developed a single-stage, transformer-based multi-scale contextual detection framework (Trans-ZSD).
• Incorporated foreground-background separation, contrastive learning for inter-class uniqueness, and commonality learning.
• Utilized balance loss to mitigate domain bias in generalized zero-shot detection (GZSD).

Main Results:

• Trans-ZSD demonstrated significant performance improvements over existing ZSD models on PASCAL VOC and MS COCO datasets.
• The framework effectively exploits inter-class correlations and optimizes feature distribution for discriminative learning.
• Single-stage approach enabled encoding of long-term dependencies and contextual features with fewer inductive biases.

Conclusions:

• Trans-ZSD offers a more robust and accurate approach to zero-shot detection by addressing key limitations of prior methods.
• The framework shows strong generalization capabilities, particularly in the challenging generalized zero-shot detection (GZSD) setting.
• Exploiting inter-class relationships and mitigating domain bias are crucial for advancing ZSD performance.