Indoor-Outdoor Image Classification
Citations: 750 | Paper Published: 03 Jan 1998 | Institutions: Martin Szummer1, Rosalind W. Picard | Authors: Martin Szummer, Rosalind W. Picard
The paper discusses how high-level scene properties can be inferred from low-level image features for indoor-outdoor scene retrieval. The proposed method achieves 90.3% correct classification on a diverse database of over 1300 consumer images provided by Kodak, outperforming state-of-the-art single-feature methods that result in about 75-86% performance.
Contributions
The contribution of this paper is to propose a different classification approach for the indoor-outdoor scene retrieval problem. The proposed method improves performance by computing features on subblocks, classifying these subblocks and then combining them in a way reminiscent of stacking. It achieves 90.3% correct classification on a diverse database of over 1300 consumer images provided by Kodak, outperforming state-of-the-art single-feature methods that result in about 75-86% performance.
Practical Implications
The practical implications of this paper are that it proposes a new method for indoor-outdoor scene retrieval, which can be used in various applications such as image search engines and autonomous vehicles. The proposed approach improves the accuracy of classification by combining low-level features computed on subblocks using stacking technique. Moreover, relatively simple classifiers like k-nearest neighbors perform better than more sophisticated neural networks and mixture of expert classifiers. This suggests that simpler methods may sometimes outperform complex ones when dealing with high-dimensional data like images.
Methods
The paper uses three low-level image features for indoor-outdoor scene retrieval: histograms in the Ohta color space, multiresolution simultaneous autoregressive model parameters and coefficients of a shift-invariant DCT. The authors systematically study these features to show how high-level scene properties can be inferred from them. They also propose computing these features on subblocks, classifying them using k-nearest neighbors classifier and then combining their results through stacking technique to improve performance. Finally, they evaluate the proposed method against state-of-the-art single-feature methods on a diverse database of over 1300 consumer images provided by Kodak.
Data
The data used in this paper is an image database consisting of 1343 consumer photographs collected and labeled by Kodak. These images depict typical family and vacation scenes, taken at different times of the year, with diverse lighting conditions such as snow, bright sun, sea, sunset etc. The dataset includes both indoor and outdoor scenes that were hand-labeled by two independent people resulting in 694 (52%) labeled as outdoor and 630 (48%) labeled as indoor.
Results
The paper achieves 90.3% correct classification on a diverse database of over 1300 consumer images provided by Kodak, outperforming state-of-the-art single-feature methods that result in about 75-86% performance. The proposed method improves the accuracy of indoor-outdoor scene retrieval problem and can be used for various applications such as image search engines and autonomous vehicles.
Limitations
- The proposed method is evaluated only on a single dataset provided by Kodak and may not generalize well to other datasets.
- Although the authors show that combining low-level features improves performance, they do not provide an in-depth analysis of why or how these features complement each other.
- The study focuses solely on indoor-outdoor scene retrieval problem and does not explore its applicability to more complex tasks such as object recognition or semantic segmentation.