Scene Categorization from Tiny Images

Humans can only just recognize images with a geometric resolution of 32× 32 pixels. However, this fact does not mean that it is also possible to assign predefined scene category labels like indoor, outdoor, kitchen or beach to such a tiny image, neither for humans nor for computers. The purpose of our work is to investigate the applicability of common scene categorization methods and the importance of color for different resolutions. We introduce a novel, challenging database of tiny images for 20 natural scene categories and present experimental results on human categorization performance. Using an algorithm based on color histograms and structure from five regions of the image, we achieve a recall rate of 46% which significantly outperforms human recall of 30% on 32×32 images. This work highlights specific properties of low resolution thumbnails and encourages future research within this area as an important preprocessing step to model context in object recognition.