Team Nishki consisted of 11th-graders is presenting a Hackathon ML solution to a Kaufland Airmap case for which they won a Datathon special award.
Used methodologies and algorithms: OCR, DarkFlow, YOLO
The solution can be found at:
https://www.datasciencesociety.net/datathon/kaufland-case-datathon-2019/
Team: Evgeni Dimov, Kalin Doichev, Kostadin Kostadinov and Aneta Tsvetkova
10. YOLO – REAL TIME OBJECT DETECTION
You Only Look Once
11. YOLO – REAL TIME OBJECT DETECTION
How does it work?
Example: Detect TV, Bicycle, Monitor
12. YOLO – REAL TIME OBJECT DETECTION
Split image into sections
Bounding boxes per section
Predictions per bounding box
Central location (within the sqare)
Width
Height
Confidence (any object)
Confidence(each class)
Remove boxes with no object
Remove redundancy – Non Max
Suppression & Intersection over Union
15. IDENTIFYING AN ISSUE
Pipeline:
Identify products & labels
Extract product numbers from labels
Apply good-old-fashioned algorithms to detect issues
Grouping into rows
Checking for unique subgroups in rows
Checking for labels with no items
Dismissing some detected objects
Running everything on the test dataset got a score of 0.677 (out of 1)