Use Open Images Dataset V7

[manifestinteractive]

Describe the Problem

Our current model is very limited on what it can detect, and also prevents us from doing several important tasks, including:

1. Detecting multiple items within the same image
2. Properly labeling the items we detected
3. Detecting more than just "fashion" items from the clothing industry

Describe the Solution

Acceptance Criteria:

• Use Open Images Dataset V7
• Make sure model used supports Mask RCNN for proper Image Segmentation
• Make sure model was trained with class names
• Ideal model will be able to be loaded by namedtuple Python library
• Ideal model should be able to be dropped into existing code

Additional Context

You can see the full level of classification this model supports in this graph:

https://storage.googleapis.com/openimages/2018_04/bbox_labels_600_hierarchy_visualizer/circle.html

From that classification tree, highest $$$ would be targeting fashion based retailers first. So stuff like belts, shoes, watches, handbags, glasses, sunglasses, etc ... all kind of grouped into apparel based stuff IMO we should target out the gate.
Kitchenware, Furniture & Toys are only other key roots classes I can see us needing for like the first year or so, making the list ( in order of priority ):

• Clothing
• Sporting Equipment
• Furniture
• Kitchenware
• Toys

However, long term goal of using this model is so we could activate these other classifications as needed.