Improved object categorization and detection using comparative object similarity

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world object distributions are long-tailed, making it difficult to gather sufficient training data for all categories.
• Effective object recognition requires sharing visual knowledge across categories to learn from limited or no examples.

Purpose of the Study:

• To leverage local object similarity information for effective knowledge transfer in object recognition.
• To develop algorithms that utilize category-dependent similarity regularization for improved learning with few or no training examples.

Main Methods:

• Developed a regularized kernel machine algorithm to train kernel classifiers for categories with scarce data.
• Adapted state-of-the-art object detectors to incorporate object similarity constraints.
• Utilized local object similarity (category pairs being similar or dissimilar) as a cue for knowledge transfer.

Main Results:

• Demonstrated significant improvements in object categorization using regularized kernel classifiers on the Labelme dataset (hundreds of categories).
• Showcased the effectiveness of category-dependent similarity regularization in enhancing object models.
• Evaluated the improved object detector on the PASCAL VOC 2007 benchmark dataset.

Conclusions:

• Local object similarity is a powerful cue for effective knowledge transfer in object recognition tasks.
• The proposed regularized kernel machine and adapted object detector significantly enhance performance, particularly for data-scarce categories.
• This approach offers a viable solution for training robust object recognizers in the presence of long-tailed data distributions.