Is Fine Grained Classification Different?

We performed experiments on two fine-grained classification tasks using a state-of-the-art pipeline (descriptor + dictionary + LLC encoding + max pooling + linear SVM). We found that this standard pipeline out-performed a dictionary-free classification technique (stacked evidence trees) that was specifically designed for fine-grained classification. The success of the method depends on two factors: (a) having high-resolution images that capture the fine detail of the objects and (b) using very large dictionaries so that this fine detail is not lost during encoding.