Smart Vending Machine.

Smart Vending Machine is a project I worked on with 5 other teammates at DeeCamp - Artificial intelligence training camp hosted by Sinovation Ventures. For this project, we collaborated with AInnovation (https://www.ainnovation.com/en) in order to overcome the challenges that exist in their smart vending machine such as difficulty in distinguishing similar goods and quickly identify new products.

THE BRIEF

The smart vending machine has become more and more popular due to its convenience and low maintenance cost. But some problems are still remained to be solved.

USER SCENARIO

SCAN FACE
Users scan their faces to unlock the smart vending machine. Their face recognition result is connected to their personal account. 
TAKE ITEMS
Once the door opens, users can take any items they need from the smart vending machine.
CHECKOUT
Once the door closes, the smart vending machine will check out the items automatically and deduct money from users' personal account. 

THE CHALLENGES

LOW ACCURACY RATE
 
The current model is more likely to make mistakes when distinguishing similar goods, the common method to overcome this problem is to avoid similar commodities appearing in the same vending machine.
NEW PRODUCTS
 
The packages are always updating and there are new products launching every day, it is impossible to retrain the model continuously to identify new products. It is very urgent to find a solution to these problems.

TECHNICAL METHODS

OBJECT DETECTION
Triplet Loss
 
The goal of triplet loss is to make sure that:
- Two examples with the same label have their embeddings close together in the embedding space.
- Two examples with different labels have their embeddings far away
FEW-SHOT CLASSIFICATION
Prototypical Network
 
For the problem of few-shot classification, learn a metric space in which classification can be performed by computing distances to prototype representations of each class. Yield substantial improvements over recent approaches
Relation Network
 
Relation Network (RN), is trained end-to-end from scratch. It learns to learn a deep distance metric to compare a small number of images within episodes. Once trained, a RN is able to classify images of new classes by computing relation scores between query images and the few examples of each new class without further updating the network
 

THE SOLUTIONS

OBJECT DETECTION
The procedure consumes a lot of manpower and material resources, what’s worse,  it relies on a lot of data. So finding a way to pre-screen similar commodities by machine is a good way to save resources and improve work efficiency. By uploading the picture of the new products, the owners of the smart vending machine can find out the similarity between existing products and new products, if they are too similar, they should not be placed together, if they are quite different, they can be placed together.
FEW-SHOT CLASSIFICATION
Few-shot learning can be implemented based on distance metric representation with only 6 samples. The owner of the smart vending machine can upload 6 pictures of the new product in different angles, the new product can then be added to the commodity library and identified by the smart vending machine.

PRODUCT DESIGN

BRAND IDENTITY
XiaoZhi - the smart vending machine. I designed the cute AI character in order to make the brand more accessible to younger audiences, who are more likely to use the smart vending machines.
LOGO DESIGN
I designed the logo for the smart vending machine app, which will be used by smart vending machine owners.

WIREFRAMES

The app allows the smart vending machine owners to add new products and identify whether a product can be placed in the smart vending machine by themselves.

EXHIBITION

Our team was selected by DeeCamp for our outstanding work to showcase in the exhibition, in which varies news media came to report the event. 
THE BOOTH
For the booth, we tried to mimic the scene of the smart vending machine. Users can take pictures and test on our app.
OUR TEAM
In the end, we took a picture with our mentor (middle) from AInnovation at the booth.
REFLECTION
I think the experience at DeeCamp was truly unforgettable, in DeeCamp I learned from world-class AI experts such as Kaifu Lee and Andrew Ng. Through working on the smart vending machine project we deepened our understanding of AI and explored the application of AI in real life. As a team, we tried to resolve the problems that exist in the smart vending machine and achieved technical innovation in similar object detection and few-shot learning for the smart vending machine. As the only designer, I was responsible for creating user-flows, designing wireframes, and producing visual assets to showcase our work, through working on this project, I learned how to create intuitive interfaces that simplify the complex AI technology in the backend and create end-to-end design solutions.

© 2020 by Queena Wang.