Dream.it : If you can dream it, you can wear it
Chat with DreamBot to build your own designs
As simple as 1, 2, 3
All your new clothing just a sketch away
We take you from sketch to dress
Search for dresses which match your schematics
*Keeps the money flowing in*
Inspiration & What it does
You're walking down the road, and see a belle rocking an exquisite one-piece. "Damn, that would look good on me (or my wife)".
You go home and try to look for it: "beautiful red dress". Google gives you 110,000,000 results in 0.54 seconds. Well that helped a lot. You think of checking the fashion websites, but the number of these e-commerce websites makes you refrain from spending more than a few hours. "This is impossible...". You perseverance only lasts so long - you give up.
Fast forward to 2017. We've got everything from Neural Forests to Adversarial Networks.
You go home to look for it: Launch Dream.it
You make a chicken-sketch of the dress - you just need to get the curves right. You select the pattern on the dress, a couple of estimates about the dress. Dream.it synthesizes elegant dresses based on your sketch. It then gives you search results from different stores based on similar dresses, and an option to get on custom made. You love the internet. You love Dream.it. Its a wonderful place to make your life wonderful.
Sketch and search for anything and everything from shoes and bracelets to dresses and jeans: all at your slightest whim. Dream.it lets you buy existing products or get a new one custom-made to fit you.
How we built it
What the user sees
Dream.it uses a website as the basic entry point into the service, which is run on a linode server. It has a chatbot interface, through which users can initially input the kind of garment they are looking for with a few details. The service gives the user examples of possible products using the Bing Search API. The voice recognition for the chatbot is created using the Bing Speech to Text API. This is classified using a multiclassifier from IBM Watson Natural Language Classifier trained on custom labelled data into the clothing / accessory category. It then opens a custom drawing board for you to sketch the contours of your clothing apparel / accessories / footwear and add color to it.
Once the sketch is finalized, the image is converted to more detailed higher resolution image using Pixel Recursive Super Resolution.
We then use Google's Label Detection Vision ML and IBM Watson's Vision APIs to generate the most relevant tags for the final synthesized design which give additional textual details for the synthesized design.
The tags, in addition to the image itself are used to scour the web for similar dresses available for purchase
Behind the scenes
We used a Deep Convolutional Generative Adversarial Network (GAN) which runs using Theano and cuDNN on CUDA. This is connected to our web service through websockets. The brush strokes from the drawing pad on the website get sent to the GAN algorithm, which sends back the synthesized fashion design to match the user's sketch.
Challenges we ran into
- Piping all the APIs together to create a seamless user experience. It took a long time to optimize the data (mpeg1) we were sending over the websocket to prevent lags and bugs.
- Running the Machine learning algorithm asynchronously on the GPU using CUDA.
- Generating a high-quality image of the synthesized design.
- Customizing Fabric.js to send data appropriately formatted to be processed by the machine learning algorithm.
Accomplishments that we're proud of
- We reverse engineered the Bing real-time Speech Recognition API to create a Node.js library. We also added support for partial audio frame streaming for voice recognition.
- We applied transfer learning from Deep Convolutional Generative Adversarial Networks and implemented constraints on its gradients and weights to customize user inputs for synthesis of fashion designs.
- Creating a Python-Node.js stack which works asynchronously with our machine learning pipeline
What we learned
This was a multi-faceted educational experience for all of us in different ways. Overall:
- We learnt to asynchronously run machine learning algorithms without threading issues.
- Setting up API calls and other infrastructure for the app to run on.
- Using the IBM Watson APIs for speech recognition and label detection for images.
- Setting up a website domain, web server, hosting a website, deploying code to a server, connecting using web-sockets.
- Using pip, npm; Using Node.js for development; Customizing fabric.js to send us custom data for image generation.
- Explored machine learning tools learnt how to utlize them most efficiently.
- Setting up CUDA, cuDNN, and Theano on an Ubuntu platform to use with ML algorithm.
What's next for Dream.it
Dream.it currently is capable of generating shoes, shirts, pants, and handbags from user sketches. We'd like to expand our training set of images and language processing to support a greater variety of clothing, materials, and other accessories. We'd like to switch to a server with GPU support to run the cuDNN-based algorithm on CUDA.
The next developmental step for Dream.it is to connect it to a 3D fabric printer which can print the designs instantly without needing the design to be sent to manufacturers. This can be supported at particular facilities in different parts of the country to enable us to be in control of the entire process.