2013 Poster Sessions : Multi-Label Classification of Large Shape Collections via Graph-Based Semi-Supervised Learning

Student Name : Qixing Huang
Advisor : Leo Guibas
Research Areas: Graphics/HCI
As larger and larger shape collections are becoming available, finding efficient ways to organize them is crucial for many applications including shape exploration, product search and data-driven shape modeling. In this paper we consider the problem of associating types and/or styles with shapes of a given category (e.g., cars, chairs, airplanes). We introduce a multi-label (i.e., a shape can have multiple labels) semi-supervised approach that takes as input a large shape collection of a given category with associated sparse and noisy labels, and outputs cleaned and complete labels for each shape. The proposed approach consists of three stages. The first stage computes a common embedding space for the input shapes, providing a powerful data structure for comparing shapes, partially or totally. The second stage jointly learns a distance function for each label class in the embedding space. These distance functions are used to construct a similarity graph over all shapes for each label class, so that shapes of this class stay close to each other in the graph. The third stage performs multi-label classification by jointly partitioning these graphs. To evaluate the performance of the proposed approach, we have created a benchmark where each shape is provided a set of ground truth labels generated by humans. Experimental results show that even with very sparse and noisy initial labels sets, the proposed approach yields results that are comparable to and at times times better than state-of-the-art supervised learning techniques requiring significantly more training data.

Qixing Huang is Post-doc researcher with Leonidas Guibas. He got his PHD degree under Prof. Guibas at Stanford in 2011. He has published papers on a variety of topics including assembling fractured objects, SLAM, matching and segmenting shapes and image editing. His recent research interest is in designing algorithms to efficiently organize gigantic shape collections.