Details

Autor(en) / Beteiligte

• Shahriari, Mana
• Bergevin, Robert

Titel

A Two-Stage Outdoor - Indoor Scene Classification Framework: Experimental Study for the Outdoor Stage

Ist Teil von

• 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), 2016, p.1-8

Ort / Verlag

IEEE

Erscheinungsjahr

2016

Quelle

IEEE Electronic Library Online

Beschreibungen/Notizen

• The state-of-the-art image classification methods require an intensive learning stage and a considerable amount of training images. Recently, with the introduction of these models (and in particular convolutional neural network (CNN)), it is believed that the best solution to achieve a system with high performance on scene classification is to learn deep scene features using CNN. While this can be true for large-scale image datasets (in the scale of one-million training images), the feasibility of reaching state-of-the-art performances in other complex datasets with a handful training examples remains unclear. Here, we argue that we can reach a highly accurate performance by presenting a simple model on the scene classification problem. This paper presents a hierarchical two-stage scene classification framework based on indoor versus outdoor notion. The suggested approach is a simple, yet a very efficient global image representation model for scene recognition. Though the term of indoor versus outdoor has been in the literature for quite a while, here we go further than that and propose a scene classifier model which classifies all the outdoor scene classes versus one undifferentiated indoor class. To achieve this, we tackle the problem of scene classification by using distributions of quantized filter responses to characterize the scenes. Later the classifier outputs either one of the several outdoor scene classes or a generic indoor scene class. We validate the proposed approach by feeding the algorithm with two benchmark datasets: 15-Scene category and SUN397. Successful experiments demonstrate the validity of the proposed approach. With the presented model, we reach an average accuracy of 97.84% and 55.1% in outdoor scenes and 93.09% and 92.4% in the indoor scene class on 15-Scene and SUN397 respectively.