AI Model For Retail Shelf Monitoring

Challenges

Grocery Store or Big marts lose revenue due to 2 reasons:

  1. Product Out of Stock: Products that are out of stock on shelves, but available in the stores is a missed opportunity. Manual process of checking stock is labour intensive and time consuming.

  2. Product Misplaced: Often misplaced product or disarranged product can cost money for business especially in high end fashion outlets where everything needs to be perfect. This links directly to customer’s sho[ping experience.

 
 
idea.png

Methodology

Created a deep learning based Model to detect out of stock or misplaced product in real time. It allowed the real time monitoring and helped in pin-pointing these issue which results in better customer experience and more business for Store or Mart. The procedure implemented is as follows:

  1. Using a CCTV, continuous video stream is getting captured.

  2. The live stream is being passed to the model.

  3. The model using that live video predicting Out of stock and misplaced items and showing them as output.

  4. System alerts like SMS/ email can be triggered alerting the right person to fix the situation.

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Original image (L) — Our model identifying misplaced products (R)

Technology Used

We used state-of-the-art learning based model and custom data-set from our nearest mart.

  1. Architecture: Retina-net with Resnet-101 as Backbone.

  2. Loss Function: Focal Loss

  3. Data-Set: Custom Data-set collected from 20 minutes video of our nearest local mart .

  4. Image Size Used: Trained on 128x128, 256x256 ,512x512 and 600x600 size images using gradual resizing to achieve higher accuracy and better generalization of the model.