Scene Categorization from Tiny Images
Andreas Wendel, Axel J. Pinz · 2007
Scene categorization can play a crucial role in applications such as image retrieval and object recognition. It seems reasonable that scene category labels like indoor, outdoor, kitchen, or beach can be assigned to images at lower spatial and radiometric resolution than may be required for categorization and localization of objects in the scenes. This paper presents a systematic investigation of resolution requirements for scene categorization. We introduce a novel, challenging database of tiny images for 20 dierent scene categories, present experimental results on human categorization performance, as well as an algorithm to eciently learn and perform scene categorization. Our model’s best recall rate of 46% significantly outperforms human recall of 30% on 32 ◊ 32 images. These results should encourage future research in the use of categorization on low resolution thumbnails as an important preprocessing step to model context in categorization, recognition and localization.