Learning object-to-class kernels for scene classification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• High-level image representations are crucial for visual recognition tasks like scene classification.
• The object bank, a representation using pretrained object detectors, has shown strong performance.
• Existing methods require improvement for greater discriminative power and semantic meaning.

Purpose of the Study:

• To propose object-to-class (O2C) distances as a novel method for modeling scene images based on the object bank.
• To develop lower-dimensional, more discriminative image representations through these O2C distances.
• To enhance the semantic interpretability of image representations derived from the object bank.

Main Methods:

• Introducing four variants of object-to-class (O2C) distances to model scene images.
• Representing images in lower-dimensional 'distance spaces' spanned by O2C distances.
• Kernelizing the distance representation to leverage the discriminant ability of O2C distances for classification.

Main Results:

• The proposed O2C distance representations achieve lower dimensionality while retaining discriminative power.
• Representations derived from O2C distances possess enhanced semantic meaning due to explicit computation.
• Extensive experiments on UIUC-Sports, Scene-15, MIT Indoor, and Caltech-101 datasets confirm significant improvements.
• The novel approaches achieve state-of-the-art performance in scene classification.

Conclusions:

• Object-to-class (O2C) distances offer a significant advancement over the original object bank approach for scene classification.
• The proposed method provides more semantically meaningful and discriminative image representations.
• Kernelized O2C distance representations lead to state-of-the-art performance on benchmark datasets.