Retail - Research Paper Reads

2 Minute Paper Summary.  Key Planogram Papers - Notes
Planogram Compliance Checking Based on Detection of Recurring Patterns

• Planogram Factors - Regulate product placements
• Conventional Planogram Compliance - Done Manually
• Compliance Checking - Detect Product layout, template matching
• Image Processing - Light conditions, viewpoints of product images, image pattern variations due to seasonal promotions

Steps

• Estimated layout vs Estimated product layout
• Spectral graph matching
• Recurring pattern detection

Conventional

• SIFT / SURF
• Cascade object detection

Recurring patterns

• Common visual pattern discovery
• Co-recognition / segmentation of objects
• Object matching with features
• Clustering of images
• Visual words and Objects

Planogram

• Number of rows in shelf
• Number of products in each row
• Type of product vs Shelf Type
• Customer brand positioning
• Competitor band positioning
• Product per customer brand influence
• Products positioned at face height has higher visibility

Metrics

• Customer presence over shelf type
• Customer presence over promotional campaign
• Customer density over shelf type
• Customer density over promotional campaign

Product Characteristics
Product Quantity, Product Size

Paper #2 - Retail Shelf Analytics Through Image Processing and Deep Learning

• Coco, Imagenet do not provide annotation to tackle instore product recognition

Recognize products

• HOG based object detection
• Viola - Jones
• SIFT
• HSV color based

Work

• Mask RCNN
• Region based detectors
• Object Detection
• Instance Segmentation
• Key point detection

For each Image

• Class of product identified
• Bounding Box position
• Number of pixels
• Number of Shelves
• Shelf Type
• Promotional Campaign type
• Cluster type

Metrics calculated on the single Picture
Shelves Number, Cluster, Shelf Type, Promotional Campaign

Implementation

• Object Detection / Mask RCNN
• Query and  find similar images against dataset
• Mask RCNN + VGG for feature extraction + KNN
• brand detector, product detector

Measure

• Predictions vs Ground truth items
• IOU

Experiments

• Per class detection
• Per class segmentation

Paper #3 - Precise Detection in Densely Packed Scenes

Key Notes

• Deep-learning based method for precise object detection
• Estimating the Jaccard index as a detection quality score
• Extensive, annotated data set, SKU-110K
• Soft Intersection over Union (Soft-IoU) network layer
• Represent detections as a Mixture of Gaussians (MoG)

Object detection

• Sliding window-based approaches
• Determine Region Proposals
• Apply Classifiers
• FPN - Feature Pyramid Network

Merge Duplicate Detections

• Non-max suppression
• Agglomerative, affinity propagation clustering

Key Experiments

• Deep IoU detection network
• Soft IOU layer
• EM-Merger unit
• Detection output (x,y,h,w)
• IOU - Intersection / Union (%% of overlap)

EM-approach for selecting detections

E Step assign each box to the nearest box cluster

Box similarity KL distance between corresponding Gaussians

Used for fast clustering

Detection Methods

• Faster-RCNN
• YOLO9000
• RetinaNet
• Infra - Intel(R) Core(TM) i7-5930K CPU@3.50GHz GeForce and a GTX Titan X GPU

Dataset - Link

Presentation - Link

Code - Link. A lot of reusable code is shared in customizing the detections

Poster - Link

Happy Learning!!!