md`# Object-Centric Social Network based on Visual Genome Dataset #### Data There are 108,077 images and 82,827 object categories in visual genome dataset. We removed objects that only appear 1 and 2 distinct images because of data scarcity, and we got 18,238 object categories. That is, 22% data over the entire data. Furthermore, we used "object alias" to merge the same objects with different alias. For example, we merge "cats" and "cat" into one category, and we got 16,800 object categories. Visual Genome: https://visualgenome.org/ #### Social Networw Construction - we build a global social network by calculating co-occurence among objects in the data. #### Individual Object-Centric Social Network - we used "spreading activiation" to extract individual social network with a energy threshold (t=0.07) from the global social network (16,800 nodes). - we identify 10 specific objects: "person", "chair", "phone", "toaster", "kettle", "bottle", "oven", "toilet", "refrigerator", "clock". #### Results object/num of images/num of neighborhood/num of distinct neighborhood
person (15909): 1568 -> 200
chair (7285): 1171 -> 205
phone (1987): 0 -> 0
toaster (184): 109 -> 29
kettle (253): 116 -> 29
bottle (4443): 747 -> 137
oven (931): 367 -> 95
toilet (2020): 631 -> 121
refrigerator (680): 356 -> 81
clock (3959): 934 -> 144
#### Visualization: Radial Tidy Tree D3’s [tree layout](https://github.com/d3/d3-hierarchy/blob/master/README.md#tree) implements the [Reingold–Tilford “tidy” algorithm](http://reingold.co/tidier-drawings.pdf) for constructing hierarchical node-link diagrams, improved to run in linear time by [Buchheim *et al.*](http://dirk.jivas.de/papers/buchheim02improving.pdf) Tidy trees are typically more compact than [cluster dendrograms](/@d3/radial-dendrogram), which place all leaves at the same level. See also the [Cartesian variant](/@d3/tidy-tree).`