Objects as Attributes for Scene Classification

Abstract

Abstract.Robust low-level image features have proven to be effective representations for a variety of high-level visual recognition tasks, such as object recognition and scene classification. But as the visual recognition tasks become more challenging, the semantic gap between low-level feature representation and the meaning of the scenes increases. In this paper, we propose to use objects as attributes of scenes for scene classification. We represent images by collecting their responses to a large number of object detectors, or "object filters". Such representation carries high-level semantic information rather than low-level image feature information, making it more suitable for high-level visual recognition tasks. Using very simple, off-the-shelf classifiers such as SVM, we show that this object-level image representation can be used effectively for high-level visual tasks such as scene classification. Our results are superior to reported state-of-the-art performance on a number of standard datasets.

Latent Pyramidal Regions for Recognizing Scenes

“…The categories are listed in decreasing order of classification accuracy of (LPR-RBF). The accuracies for (KSPM) and (OB) are taken from [17]. By comparing the per-class accuracy of our LPR representation with OB [17] and KSPM [7] which is summarised in Table 4, we see that the accuracy of the proposed method is significantly higher than the other two in most of the categories.…”

“…The accuracies for (KSPM) and (OB) are taken from [17]. By comparing the per-class accuracy of our LPR representation with OB [17] and KSPM [7] which is summarised in Table 4, we see that the accuracy of the proposed method is significantly higher than the other two in most of the categories. This suggests that the proposed method is less confused by similar backgrounds such as sailing and rowing and similar visual structures like human players in polo and bocce.…”

“…In [16,17], Li et al propose another unique object-centered approach called the Object Bank(OB). In this model, each image is represented as the response map of a large number of pre-trained generic object detectors.…”

“…GIST [21] and KSPM [7], which capture the global scene properties) and high-level semantical image representation methods (i.e. MM-Scene [22] and OB [17,16]). Furthermore, by using an SVM with radial basis-function kernel for classifying we obtain an accuracy of 86.25% which is a more than 4% increase over our baseline.…”

“…Here, the linear kernel did not result in a significant improvement over LPR-MS but it's accuracy is still significantly better than LLC. [21] 63.88 OB1 [17] 77.88 MM-Scene [22] 71.7 LLC(our global term) 81.87 KSPM [7] 71.57 LPR-MS(our approach) 85.0 WWW [18] 73.4 LPR-LIN(our approach) 85.2 OB2 [16] 76.3 LPR-RBF(our approach) 86.25 Table 4. Per-class accuracy of the proposed method (LPR-MS and LPR-RBF) compared with our baseline (LLC [4]) and state-of-the-arts (KSPM and OB) on UIUCSports dataset.…”