Computer Vision for Retail

Industry: Retail

Introduction

Client is a leading retail store from Hong Kong. Client’s stores offer a selective range of daily consumables including packaged drinks, candies, snacks, newspapers and magazines, etc. The stores also provide a range of convenience services such as bill payment, reloading, ticketing and cash withdrawal services. To further enhance the range of convenience services, client has launched e-fulfilment services for leading e-commerce websites including Finger shopping, Dimbuy etc.

Problem Statement

  • There are approximately 350 client’s retail stores in Hong Kong. Peak hour sees significant large number of footfall and long queues. Check out time could be a couple of minutes for customers at the back of the queue.
  • Client was looking for some innovative solution to reduce the checkout time for their customers.
  • Abzooba aligned with the Client’s motto of – “Let’s make it easy” and suggested a contactless checkout experience using Vision Computing technologies.
  • Abzooba also suggested several possible avenues to improve customer check out experience within the store.

Project Journey

What Abzooba offered to Client

Objective:

Objective of the solution offered to the client was to perform a small proof of concept by identifying 10 SKU’s from their store with the help of Computer vision technology. Abzooba developed and implemented contact less easy checkout solution by correctly identifying the chosen 10 SKU’s in initial phase through a setup consisting of a camera, tripod and displayed the barcode on display screen.

Approach:

Abzooba implemented end to end flow of contact less checkout solution by following steps:

Step 1: Instance detection for easy checkout
  • Setting up the required Hardware: Camera/mobile camera positioned on top of a tripod at a certain angle to have a 2.5ft by 2.5ft white table top surface.
  • Data Generation:
    a. SKU Images of a product/SKU from all possible view angles and orientations.
    b. Scene Images: of a real/synthetic scenario – multiple products on table
  • Labeling: Have SKU/instance specific bounding box annotations on the captured
    (and/or synthetic) scene images
  • Training: Learn a deep neural net to detect the correct SKU instances from a scene
  • Inference model: Publish the best performing model on cloud/premise/device
Step 2: EPOS Integration with camera/hardware setup
  • Abzooba developed Client application on EPOS device to:
    • Get input feed from camera device to capture an image of products during checkout
    • Communicate with ML API engine
  • UPLOAD image captured via POST request
    • Get JSON response – list of SKU/barcodes
    • Communicate with billing software
  • Share barcodes/SKU info to generate bill on the billing software

Business benefits:

  • Abzooba developed Computer vision object detection model with 93% mean precision
  • Initial checkout time at Client’s store was around ~6 seconds. After Abzooba’s contact less checkout solution, checkout time was reduced to ~1.5 seconds.