Continuity Report: Revisiting Grocery Recognition using TensorFlow Samuel C. Hoffman, Dilip Thiagarajan {sch235, dt372}@cornell.edu Last Updated: September 15, 2016 1 Introduction The revolution of convolutional neural networks (ConvNets) in the field of machine learning has spawned applications in a wide array of areas. Some accessibility applications have already begun to harness the potential of ConvNets in computer vision for assisting the visually impaired such as Facebook’s Automatic Alternative Text tool for narrating images [4] However, the underlying tech- nology involved in recognizing grocery items to assist visually impaired shoppers has not advanced much beyond the study by Merler, Galleguillos, and Belongie [10]. We aim to revisit this study and apply modern deep convolutional neural networks to solve the problem. We will use the open source library, TensorFlow 1 , to implement the deep learning system [1]. We attempt to solve the problem of recognizing objects in the wild using images of the object taken under ideal conditions to train. This challenge has enumerable applications but in this study, we will look specifically at recognizing grocery products sold in Mattin’s, a local café on the Cornell campus in Ithaca. However, the results should be easily transferrable to other areas of application. Ideally, training data and testing data would be obtained from the same distribution, however, in real-world applications, it is often more convenient to obtain training data from the web or dedicated datasets. Ultimately, our goal in this study is to achieve real-time recognition and localization with a camera. The GroZi project 2 , from which the original study came, encompasses many applications sur- rounding this core recognition problem. However, the implementation still relies on color histogram matching, SIFT matching, and boosted Haar-like features and the results left room for improvement. For further background, please read our original proposal. 3 1 https://www.tensorflow.org/ 2 http://grozi.calit2.net/ 3 https://github.com/dthiagarajan/grozi_tf/blob/master/proposal/Proposal.pdf 1