Dataset

So far this data set is build on :

1. Well-know dataset with object detection annotation :

• Google OpenImage
• ImageNet (Standford)
• VisualGenome

2. Other image databases, which will require human annotation (TBD)

As of today we have identified the following amount of annotations within these providers :

classes / provider	visualgenome	openimage	imagenet	total
hammer	35	139	427	601
screwdriver	29	85	479	593
paintbrush	7	0	533	540
drill	0	203	421	624
wheelbarrow	0	0	543	543
	94	427	2403

file classes2datasources_synsets.json

This file define the categories classes we wants to target in each datasource provider. So it contains a map of provider, and a dictionnary of common classes, with a mapping to each provider own class name