VLDUS: Vision-language distillated unseen synthesizer for zero-shot object detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative methods in zero-shot object detection (ZSD) synthesize features for unseen classes using semantic embeddings.
• Current methods suffer from poor diversity and generalization due to reliance on limited seen-class data for training feature synthesizers.

Purpose of the Study:

• To develop a novel knowledge distillation-based feature generation paradigm for ZSD.
• To enhance the diversity and generalization ability of synthesized unseen class features.

Main Methods:

• Proposed the Vision-Language Distillated Unseen Synthesizer (VLDUS) for ZSD.
• Implemented two complementary generative distillation strategies to transfer knowledge from a pre-trained CLIP model.
• Employed feature-aligned and relation-aligned generative distillation to improve generalization and intra-class diversity.

Main Results:

• VLDUS generates unseen features with high intra-class diversity and inter-class separability.
• Achieved significant performance improvements on ZSD and Generalized ZSD (GZSD) tasks.
• Outperformed state-of-the-art methods on benchmark datasets including MS COCO, PASCAL VOC, and DIOR.

Conclusions:

• VLDUS effectively addresses limitations in current ZSD generative methods.
• The proposed distillation strategies enhance the quality and applicability of synthesized features.
• VLDUS offers a robust solution for improved zero-shot object detection capabilities.