Description
At Fieldguide we are developing a digital field guide for all species of flora and fauna across the planet. We are using image recognition technology to enable species identification and to help with the curation of this massive catalogue. In this talk I will describe how we are building an image recognition system with the aim of identifying all known species in the natural world. The system has gone through a number of iterations at this point using a variety of computer vision and machine learning techniques, from nearest neighbour search to classification with fine tuned deep convolutional networks. All of this has been implemented in Python using scikit-learn, Numpy, Caffe and Tensorflow. Aside from the obvious machine learning challenges in designing and training such a system we faced numerous technical challenges while implementing and scaling this system in a cost effective manner. I will discuss these challenges, our solutions and the remaining open problems. While this talk will be relatively high level with few code examples and no math (but lots moths), it will be of most interest to those who have some knowledge of machine learning concepts.