Object Discovery: Soft Attributed Graph Mining

Object Discovery: Soft Attributed Graph Mining

Quanshi Zhang, Xuan Song, Xiaowei Shao, Huijing Zhao, and Ryosuke Shibasaki
in IEEE Trans on PAMI 38(3):532-545, 2016

Conference version: Attributed Graph Mining and Matching: An Attempt to Define and Extract Soft Attributed Patterns, in CVPR 2014.

You can download the paper, the code, and the RGBD image dataset, and the web image dataset.

Given an initial graph template and a set of attributed relational graphs (ARGs), this method modifies the graph template into the common subgraph pattern among the ARGs with the MAXIMAL graph size, by discovering probably missing nodes, deleting redundant nodes, and training the attributes.


Extented Application 1: As a platform of category modeling from cluttered scenes, this technique has been applied to train category models for 3D reconstruction from ubiquitous images.

Quanshi Zhang, Xuan Song, Xiaowei Shao, Huijing Zhao, Ryosuke Shibasaki, “When 3D Reconstruction Meets Ubiquitous RGB-D Images”, in CVPR 2014 (PDF)


Extented Application 2: Recovering the model for the whole object from a fragment using cluttered (web) images.

Given a set of cluttered scenes (web images) that contain objects in the target category, this technique can be applied to recover the category model from “an object fragment”.

Extented Application 3: Mine a deformable model from unlabeled videos for tracking and pose estimation of animals.

Web image dataset (download) for visual mining


Please contact Dr. Quanshi Zhang, if you have questions.