Furniture recognition/detection using a fine-grained multi-label Conv-Net/Caps-Net subcategory classification: a research on detecting multiple attributes of a furniture-object i.e. category, color, material, shape, and style - BC-401

Preferred Disciplines: Machine Learning and Artificial Intelligence Scientist/Engineer (Masters or PhD)
Project length: 4-6 months (1 unit)
Approx. start date: As soon as possible
Location: Burnaby, BC
No. of Positions: 1
Preferences: Hands on student with strong coding background in Python on both Windows and Linux OS and preferably have experience with production and industry-level coding work.
Company: Cymax Incorporation

About Company:

Cymax is one of the fastest growing online furniture retailers with annual sales exceeding US $100 million. With over 75,000 SKUs, Cymax is a leader in online sales for all items home and office. Internet Retailer Magazine ranked Cymax within the Top 200 e-tailers in the world in both 2011 and 2012.

Recently, Cymax is incorporating data-driven solutions to increase sales and margine. Therefore, Cymax is incorporating cutting-edge machine learnig and AI algorithms into various projects ranging from pattern recognition in big-data to computer vision and natural language processing. As a result, Cymax provides “real-world” big-data and supports research and development in the field of Artificial Inteligence to advance the field with the focus on supporting scientists in the fields of supervised and reinforcement learning. 

Summary of Project:

Designing an automated recommender system to identify the trending products on Amazon pltform and finding the most visually-similar products in our database has been our challenge and interest for the past couple of years. To make this happen, we need to develop a data-driven computer vision program to help us identify different categories and sub-categories of the furnitures as well as different attributes such as color, shape, style, and material to increase the accuracy of the final search. Since identifying different categoiries of furnitures only based upon the image is a very challenging task due to the very close similarity of the different categories, a rigorous research must be done to overcome this problem to make our current visual search more accurate. Once this is done, we are planning to use the same trained algorithm for other projects such as “furniture detection”, “interior design recommender system”, “recommender bought-together items based on visual similarity”, “automatic title/description generator for the home décor images” and many other areas in our business.

Research Objectives/Sub-Objectives:

  • Coming up with a novel classification method to identify “highly similar” subcategories from each other.
  • Novel methods to identify the features of the furnitures in an image such as color, material, texture, specific visual patterns, style, shape and so on.
  • Coming up with methods to make the final vision algorithm work for the “real-world” data (not only for the synthesised data).

Methodology:

    • Fine-grained multi-labled CNN model for subcategory classification.
    • Using novel methods to capture the objects dominant color, texture, material, and visual patterns to make a better and faster visual search.
    • Using other novel methods such as Visual Attention models such as Visual Autoencoders or novel nueral net architechtures such as Capsule Networks to increase the accuracy of the visual classification and search. 

    Expertise and Skills Needed:

    • Strong coding background in Python.
    • Experienced with deep-learning libraries such as TensorFlow, Keras, and PyTorch and also RL platforms such as DQN and Google’s Dopamine.
    • Interested in applying mathematical models to solve industrial DL problems.
    • Familiar with fundamentals of supervised-learning, unsupervised-learning, and reinforcement-learning  algorithms.
    • Familiar with how deep learning algorithms work. Familiar with the fandamentals of for example CNN and RNN algorithms and how they recognise patterns in the data.
    • Familiar with Generative and Adverserial algorithms and knowing how to implement and apply them to approach a problem.   
    • Critical thinking and having a can-do attitude.    

    For more info or to apply to this applied research position, please

    1. Check your eligibility and find more information about open projects.
    2. Interested students need to get the approval from their supervisor and send their CV along with a link to their supervisor’s university webpage by applying through the webform
    Program: