Collaborative Representation based Fine-grained Species Recognition

Fine-grained Visual Categorization (FGVC) is an open problem in Computer Vision due to subtle differences between categories. The present paper demonstrates that Collaborative Representation based Classification (CRC) can address this problem successfully. Instead of the traditional discriminative approach of classification, CRC takes a co-operative approach by representing the query image as a weighted collaboration of training images across all classes in the feature space. The superior performance of CRC compared to some other modern classifiers including SVM is shown in this work using several popular descriptors like GIST+Color, SIFT and CNN features with Species Recognition chosen as the representative FGVC problem. Besides experiments on the Oxford 102 Flowers and CUB200-2011 Bird benchmarks, the present work also introduces a new challenging dataset NZBirds v1.0 with 600 images of 30 New Zealand endemic and native bird species.