We crawled Amazon, Crate and Barrel ,
Car Dimensions, Dimensions, Wikipedia,
and Google for raw size data. After removing categories
without any size information and merging the similar categories, we build a dataset of object sizes with
a five-level tree structure based on WordNet with about 800 categories. This dataset is general
and covers most common semantic objects in ordinary indoor and outdoor scenes.
This is the structure of Metric-Tree:
Due to copyright,we will not release the images, but we provide the image links with the size dataset, so you can download the images by yourself. There are the Metric-Tree and Metric-Tree Size Subset (Which we used in our expriments).
You can use the code to build the basic GMM for each categories and download the images.
├── readme.md: the description of Metric Tree and Code
├── meta.json: the organization of Metric Tree
├── code: the code tools
└── metrictree: the size data and annotations
meta.json: the node information of Metric-Tree,each node (which indicates a category) has name,children and values (the number of the catgeroy items).
code: it includes the code of downloading images and building the GMM model for each category.
metrictree: It's include the size data, each folder represents a category, and each folder includes the items of size data and annotations.
unit of size: millimeter.
Usually, most items include width,depth and height, and if there is no dimension,such as no height,we will use None for it.
We select 43 categories (which has sharp dimensions) to build the subset for our experiments. This subset only include sizes. We merge some categories,remove the None dimension and fix some dimension errors, so some of the categories could be a little different from Metric-Tree.