Towards generic image classification using tree-based learning: An extensive empirical study

• Random subwindow extraction is able to capture many different types of image patterns.
• Learning features with trees yield better performances than direct classification.
• Supervised feature learning from raw subwindow pixels is better than unsupervised.
• Simple optimizations can increase significantly recognition performances.

This paper considers the general problem of image classification without using any prior knowledge about image classes. We study variants of a method based on supervised learning whose common steps are the extraction of random subwindows described by raw pixel intensity values and the use of ensemble of extremely randomized trees to directly classify images or to learn image features. The influence of method parameters and variants is thoroughly evaluated so as to provide baselines and guidelines for future studies. Detailed results are provided on 80 publicly available datasets that depict very diverse types of images (more than 3800 image classes and over 1.5 million images).