Influence of the amount of context learned for improving object classification when simultaneously learning object and contextual cues

Abstract

Humans use visual context to improve object recognition. Yet, many machine vision algorithms still focus on local object features, discarding surrounding features as unwanted clutter. Here we study the impact of learning contextual cues while training an object classifier. In a new image database with 10 object categories and 28,800 images, objects were presented in contextual or uniform backgrounds. Both the fraction of contextual backgrounds during training and the spatial extent of context were analysed. Local object features and broader context features were extracted by two biologically inspired algorithms, previously used for object and scene classification, respectively: HMAX, applied to a tight window around every object, and a “Gist” algorithm, applied to a larger yet still localized window. The descriptors from both algorithms were combined and processed by a Support Vector Machine. The recognition rate increased from 29%, without contextual cues, to 43% for objects presented in their context.

Keywords:

• Amount and size of context
• Biologically inspired algorithms
• Learning
• Local context
• Object classification

Acknowledgements

Supported by the National Science Foundation (grant number BCS-0827764), and the Army Research Office (W911NF-08-1-0360 and W911NF-11-1-0046), the Defense Advanced Research Projects Agency (HR0011-10-C-0034), and U.S. Army (W81XWH-10-2-0076). The authors affirm that the views expressed herein are solely their own, and do not represent the views of the United States government or any agency thereof.

Notes

¹ www.ilab.usc.edu

²Note that for p_context=0% and p_context=100% the whole corresponding pool of training image is used resulting in 10 identical selections.