This is a general overview of MSCV capstone project sponsored by Bossa Nova Robotics and advised by Prof. Katerina Fragkiadaki. Team members include Xunyu Lin and Zheyu Zhang, both from MSCV class of 2018.
Supermarket employees have to deal with inventory problems, such as out of stock, product misplacement, etc. Bossa Nova Robotics (BNR) is the leading developer of robots designed to provide real-time inventory data for the global retail industry. To achieve this, BNR builds robots that can navigate retail aisles autonomously and capture high resolution images of every product on every shelf.
The objective for our capstone project is to learn a good feature embedding space of product facings. A good embedding space with similar products clustered closer and other products further is beneficial for future tasks like product association, grouping and recognition.
In our project, we explored the possibility of modeling such embedding space with similarity learning. We first experimented on different fully-supervised similarity learning methods with different hard mining techniques. Further, we proposed semi-supervised learning methods as well as active labeling scheme to alleviate the burden of human labeling for training. Our final semi-supervised method reached better performance than fully-supervised method. (Details under NDA).