Semantics-Guided Contrastive Network for Zero-Shot Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot object detection (ZSD) aims to detect objects from unseen categories, extending conventional detection models.
• Existing ZSD methods often suffer from bias towards seen categories and poorly structured visual feature spaces.
• This limits their ability to effectively generalize to novel object classes.

Purpose of the Study:

• To address the limitations of current ZSD approaches by developing a novel framework.
• To improve the learning process by incorporating semantic information of unseen classes.
• To enhance the discriminative power of visual features for better visual-semantic alignment.

Main Methods:

• A novel Semantics-Guided Contrastive Network for ZSD, named ContrastZSD, is proposed.
• ContrastZSD utilizes two semantics-guided contrastive learning subnets for region-category and region-region pair contrasts.
• Supervised contrastive learning leverages ground truth labels and class similarity for training.

Main Results:

• ContrastZSD effectively learns more knowledge about unseen categories, mitigating bias towards seen concepts.
• The framework optimizes the visual data structure, making it more discriminative for ZSD.
• Experiments on PASCAL VOC and MS COCO benchmarks demonstrate superior performance over state-of-the-art methods.

Conclusions:

• ContrastZSD represents a significant advancement in zero-shot object detection.
• The proposed contrastive learning approach effectively tackles key challenges in ZSD.
• The method shows strong potential for real-world applications requiring detection of novel object categories.