Attribute Predictions for Flowers

In this work, we apply and advance deep learning methods for computer vision in the context of a challenging flower dataset. Flower species are often differentiated only by very subtle indicators. However, images of a single species may still vary hugely in pose and scale, each image revealing only a subset of relevant distinguishing attributes. Experts or hobbyists will often use botanical keys to identify unknown flowers by examining various such attributes, e.g. colour or symmetry (which may each be shared across many species), or by placing the flower in a broader taxonomic group like a botanical family. Even for an expert human observer, precise classification of a flower’s species from a single image may be impossible. Our aim is therefore to acquire as much information as possible from a single flower image, by predicting the values of botanical attributes and taxonomic classifications. Further, we wish to provide useful information about a flower even when training data for its species are limited or nonexistent. In this section, we will introduce the datasets used, attributes studied, and primary challenges to working with these botanical images and labels. Section 2 will discuss classification at the species level, demonstrating improvements on the state-of-the-art for the OF261 dataset. In section 3 we will introduce and compare a variety of loss functions suited for training models to predict attributes with ambiguous multiple-label sets. Section 4 will evaluate these approaches to attribute predictions with results on synthetic data as well as the flower images. Finally, section 5 will explore directions for future work, discussing zero-shot learning and embeddings.