Vocabulary-Informed Zero-Shot and Open-Set Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object categorization faces challenges with limited labeled data and large, open-set label recognition.
• Existing zero-shot learning methods have limitations with vocabulary size and class separation.

Purpose of the Study:

• To propose a unified framework, vocabulary-informed learning, to address challenges in object categorization.
• To improve performance in supervised, zero-shot, generalized zero-shot, and open-set recognition.

Main Methods:

• Introduced a weighted maximum margin framework for semantic manifold-based recognition.
• Incorporated distance constraints from supervised and unsupervised vocabulary atoms.
• Ensured labeled samples project closer to correct prototypes in the embedding space.

Main Results:

• Demonstrated improvements across supervised, zero-shot, and generalized zero-shot recognition tasks.
• Achieved success in large open-set recognition with up to 310K class vocabulary.
• Validated performance on Animal with Attributes and ImageNet datasets.

Conclusions:

• Vocabulary-informed learning offers a unified approach to various recognition problems.
• The proposed framework effectively handles large-scale and open-set recognition challenges.
• This method advances object categorization by leveraging comprehensive vocabulary information.