Fine-grained image classification with object-part model

Fine-grained image classification is used to identify dozens or hundreds of subcategory images which are classified in a same large category. This task is challenging due to the subtle inter-class visual differences. Most existing methods try to locate discriminative regions or parts of objects to develop an effective classifier. However, there are two main limitations: (1) part annotations or attribute descriptions are usually labor-intensive, and (2) it is less effective to find spatial relationship between the object and its parts. To alleviate these problems, we propose a novel object-part model that relies on an attention mechanism. The main improvements of our method are threefold: (1) an object-part spatial constraint which selects highly representative parts, able to keep parts both discriminative and integrative, (2) a novel heatmap generation method, able to represent comprehensively the discriminative parts by regions, and (3) a speed up of the part selection by filtering image patch candidates using a fine-tuned CNN. With these improvements, the proposed method achieves encouraging results compared to the state-of-the-art methods benchmarking on the Stanford Cars and Oxford-IIIT Pet datasets.

Duke Scholars

Author Kaizhu Huang DKU Faculty