Technical Briefs

Methods and Applications

Saturday, 01 December 11:00 - 12:45 |  Garnet 219

Scoring functions for Automatic Arrangement of Business Interiors - Picture

Scoring functions for Automatic Arrangement of Business Interiors

Automatic interior arrangement problem is usually described as either an optimization or a procedural generation task. We introduce a new method of addressing this problem by means of scoring functions. Our system creates high quality and diversified arrangements within tight time constraints.

Szymon Chojnacki, Reterio Project, Poland


Real-time Manga-Like Depiction Based on Interpretation of Bodily Movements by Using Kinect  - Picture

Real-time Manga-Like Depiction Based on Interpretation of Bodily Movements by Using Kinect

In order to enrich visual communications, we propose a real-time system that interprets bodily movements by using Kinect. Based on captured data analysis, our system automatically attaches various Manga-like effects, such as speed lines, focus lines, or motion blur, to live-action movie in real time.

Daiki Umeda, Tokyo Denki University Graduate School
Moriya Tomoaki, Tokyo Denki University Graduate School
Tokiichiro Takahashi, Tokyo Denki University Graduate School


Automatic Chinese Food Identification and Quantity Estimation - Picture

Automatic Chinese Food Identification and Quantity Estimation

Our system can automatically identify food categories and has been implemented as an Andriod application. The overall accuracy for 50 categories of food achieves 68.3% by cooperating with SVM and multi-class Adaboost algorithm. Top 3 and Top 5 candidates accuracy can reach 84.8% and 90.9%.

Mei-Yun Chen, National Taiwan University
Yung-Hsiang Yang, National Taiwan University
Chia-Ju Ho, National Taiwan University
Shih-Han Wang, National Taiwan University
Shane-Ming Liu, National Taiwan University
Che-Hua Yeh, National Taiwan University
Ming Ouhyoung, National Taiwan University


Using Text N-Grams for Model Suggestions in 3D Scenes - Picture

Using Text N-Grams for Model Suggestions in 3D Scenes

We suggest models for a partially completed 3D scene by examining label co-occurrences in the Google Web 1T 5-gram dataset. Our new text-based algorithm injects a greater variety of good model arrangements into a recent Graph Kernel system that only trains on available 3D scene data.

Laureen Lam, Stanford University
Sharon Lin, Stanford University
Pat Hanrahan, Stanford University