vg_cleasing

dataset cleansing for Visual Genome

Introduction

Since the relationship dataset from Visual Genome are extracted from sentences, they look a little messy. Therefore, to make full use of the relationship dataset, I do some dataset cleansing beforehand.

Some intermediate analyzing results are stored in Google Sheet. Feel free to check it.

All the preprocessing are implemented with Python. For easy interaction, I also use iPython notebook.

For more information, don't hesitate to contact me.

File organization

For further convinience, please organize the dataset as following:

- ROOT_PATH (the root dir for the dataset)
	- VG_100K_images (images)
	- vg_cleansing (the python scripts for visual genome cleansing)
		- temp_data (to store some temporary data for dataset cleansing, e.g. object categories of Pascal VOC)
		- models ([optinal] to store the word2vec models)
	- Annotations (annotations for visual relationship)
	- cleansed_data (the output of our cleansing)