Improving object classification by simultaneously learning object and contextual cues

摘要

Numerous studies have demonstrated the use of contextual cues to improve object classification by human viewers. Inspired by human perception, a growing number of studies investigate the role of context, previously seen as clutter, for object classification. We investigate the impact of learning contextual cues while training an object classifier. Object and context features were extracted by two algorithms: Hmax (Serre et al. IEEE PAMI, 2007, 29(3):411-426) and a gist algorithm (Siagian et al. IEEE PAMI, 2007, 29(2):300-312), usually used respectively for local object classification and global scene classification. These different features were then combined into a single vector and processed by a support vector machine (SVM) for learning an object classifier. The influence of context on classification learning is studied using a new image database with 5 object classes in consistent and in random contexts (total 1,000 images). The influence of both the spatial extent of a context window around the object and the fraction of consistent contextual exemplars vs. random exemplars were analyzed. Increasing the number of consistent exemplars improved classification when objects were presented in their consistent context but penalized it when objects were in random context. A tighter contextual window was also more helpful than a wider one. Combining the features of the objects with the features of a spatially limited window around the object improved the classification compared to using object features alone (The average over the 5 object classes of true positive rate was 93% when Hmax and the gist algorithm were combined vs 82% when Hmax was used alone. These results were obtained using training and testing images that all contain objects in consistent context). Our results show quantitatively that consistent context is helpful for object classification.