Graduate School of Engineering, University of Yamanashi, Yamanashi, Japan.
|
Ohbuchi Laboratory Graduate School of Engineering, University of Yamanashi, Yamanashi, Japan. |
 |
大渕 竜太郎
Ryutarou Ohbuchi Professor Computer Science Department University of Yamanashi 4-3-11 Takeda, Kofu-shi, Yamanashi-ken 400-8511, Japan ohbuchi *a t* yamanashi d0t ac d0t jp |
 Nishi-adzuma peak, Fukushima, Japan. |
Ryutarou Ohbuchi
Professor
Computer Science Department, University of Yamanashi
I teach at the Computer Science Department (or, Computer and Media Engineering Department) of the faculty of engineering, University of Yamanashi.
Ohbuchi Lab. Internal Page.
My publication list and resume
Research
My Research
講義 Courses
(これらのページは更新され，授業やその課題，試験などの情報が随時追加変更される．授業を受けている学生は頻繁にチェックすること．.)
2009年度 前期
2009年度 後期
Recent News
2009年度
Specification of a query is one of the most fundamental issues in retrieving multimedia objects such as images and 3D geometric models. While query by 3D model example (or 2D sketch example) of a desired shape is quite effective, text based search and retrieval of 3D models, or Query By Text (QBT) approach, has its set of advantages. For example, specifying semantics or intention may be easier than by query by shape example. An issue in the QBT approach is the lack of labeled 3D models. This paper discusses a method to add text tags to 3D models without tags. Given a set of 3D models with tags, that is, the labeled training set, the method propagate the likelihoods, or Tag Relevance Ranks, of the tags by means of multiple overlapped manifold ranking [Zhou03] to the other models.
Ryutarou Ohbuchi and Shun Kawamura, Shape-Based Autotagging of 3D Models for Retrieval, accepted, 4th International Conference on Semantic and Digital Media Technologies (SAMT 2009), Graz, Austria, Dec. 2-4, 2009. (PDF)
This paper describes a method for searching full 3D models from a range scanned 3D mesh from a viewpoint. The method is bases on the 3D model retrieval algorithm based on local visual features we have publised previously (Ohbuchi_SMI08). However, significant changes are made to deal with issues associated with range scanning, such as cracks and jagged edges of the query 3D mesh. Those changes include dense and random placement of Lowe's Scale Invariant Feature Transform (SIFT) sample points, as well as an importance sampling of lower-frequency, i.e., larger-scale, images of the Gaussian image pyramid used in the SIFT. Experimental evaluation showed that the method significantly ourperformed the methods in SHREC 2009 Partial View Retrieval track (organized by Afzal Godil, et al.); the proposed method scored at Mean First Tier = 37%, whereas the two entries to the SHREC scored 15% less, at Mean First Tier = 22%.
We have improved execution speed and retrieval performances of the 3D model retrieval algorithm based on local visual features we have publised previously (Ohbuchi_SMI08). The method employs dense random sampling of SIFT features in the multi-view images to capture more features than the original, salient-point based method. To compensate for the increased cost of computation due to much larger number of local visual features, we adopted both GPU based feature extraction [Ohbuchi_S-3D_08] and a fast randomized decision tree algorithm for codebook learning and visual word encoding.
We accelerated some of the steps for the Bag-of-Features SIFT algorithm for 3D model retrieval (Ohbuchi_SMI08) we have published at the SMI 2008. We adopted Wu's algorithm for GPU-based SIFT feature extraction and table lookup for distance computation for a significant speedup without degradation in retrieval performance.
Shape-based comparisons of 3D models is affected by shapes as well as semantics of the 3D models. The semantics may be classified into long-term, well-established, share semantics and per-query-session semantics (or intention). Previously, most of so-called "semantic" retrieval algorithms took only one of these types of semantics into consideration. Our novel approach combines a set of long-term, established semantic classes with per-query-session semantics for 3D model retrieval. Our method first learns, off-line, a set of multiple semantic classes by using our semi-supervised dimension reduction approach Ohbuchi, MIR07] to produce a "semantic manifold" from the input, ambient features. The method then applies manifold ranking-based relevance feedback (RF) on the semantic manifold. Our experimental evaluation showed that the RF using manifold ranking performed on the semantic manifold far outpeforms the same applied to the original ambient feature space.
We won the two tracks in the SHape REtrieval Contest (SHREC) 2008, the CAD Model Track 2008 and Generic Models Track. We entered the two tracks using our three algorithms.
CAD Model Track (https://engineering.purdue.edu/PRECISE/shrec08)
We won the CAD Model Track 2 consecutive years.
U_Yama_A won with a very large margin as the graph shows. For the details, see the SHREC 2008 CAD track page.
U_Yama_A
U_Yama_B*
* Two methods named "U_Yama_B" that appear in the Generic models track and the CAD Model Track are different. The U_Yama_A methods in the two tracks, on the other hand, are the the same.
U_Yama_A (the same method as the U_Yama_A) won the query set 1.
U_Yama_B (different method from the U_Yama_B for the CAD model track) won the query set 2.
U_Yama_A
U_Yama_B*
* Two methods named "U_Yama_B" that appear in the Generic models track and the CAD Model Track are different. The U_Yama_A methods in the two tracks, on the other hand, are the the same.
We have developed a method to compare 3D models by their shape using a set of local visual features. Our paper has been accepted as a regular paper for the SMI 2008. By using local visual features, the method performs very well in comparing shapes of articulated or deformable shapes, such as human beings, snakes, insects, etc. Evaluated by using the McGill 3D Shape Benchmark for the articulated figures, it far outperformedall the global shape features we tested.
For the rigid shapes, evaluated by using the Princeton Shape Benchmark, the method performs about as well as some of the most powerful method employing global shape features.
Similarities among 3D mordel "shapes" are often influenced by their semantics as much as their shape. Semantic influcence may be captured by on-line learning via relevance feedback, or by off-line learning of semantic categories in a training database.
In this study, we have used a semi-supervised off-line learning of semantic categories to boost 3D model retrieval performance. The method and experimental evaluation results are described in the paper below. It is a draft of a paper accepted as an oral paper for the 9th ACM SIGMM International Workshop on Multimedia Information Retrieval (ACM MIR 2007) to be held in Augsburg, Germany. This research a part of Akihiro Yamamoto's Master's thesis work.
半教師付き学習をつかって意味カテゴリを学習することにより，3次元モデルデータベースの検索性能を上げることに成功しました．この手法は，9th ACM SIGMM International Workshop on Multimedia Information Retrieval (ACM MIR 2007)の口頭発表として採録された論文 (PDF)に述べてあります．
Performance of this 3D model retrieval method was measured using SHREC 2006, and is listed here.
The new method outperforms our previous method by about 15% in terms of Mean First Tier (Highly relevant) measure (57.9% for the new method), or the Mean Average Precision measure (0.618 for the new method).
Our new 3D model retrieval method (Method_A in the list) that employes off-line supervised learning of semantic categories is evaluated by using SHREC 2006 benchmark.It outperforms our previous method (based on unsupervised learning) as well as all the methods that entered the SHREC 2006 by significant margin.
3次元CADモデル検索コンテスト（SHREC2007 CADトラック）で我々が優勝！
We won the SHREC 2007 (Shape Retrieval Contest) CAD model track. Our method won the competition by a significant margin, as the result shows.
The method we have employed is described in this short paper for the SHREC 2007 report. The contest's results are presented at a speciall session at the Shape Modeling International 2007 (SMI 2007,) in Lyon, France.
We basically used the feature dimension reduction based on unsupervised learning described in the papers (ACM MIR 2006, R. Ohbuchi, J. Kobayashi, PDF) and (AMR 2007, R. Ohbuchi, J. Kobayashi, A. Yamamoto, PDF).
We also used the multiresolution shape comparison approach we have proviously proposed (R. Ohbuchi, T. Takei, PG 2003, PDF).
It should be noted that we did not use "CAD models" for the learning; we used 5,000 "generic" 3D models to train the Locally Linear Embedding algorithm by Roweis et al.
SHREC 2007 ３Dモデル検索コンテストの CADモデル分野で1位になりました〜！順位はここを見てください．
手法はこの論文に書かれています．教師無し学習を用いて多数の3次元モデルの特徴から，これら特徴の張る部分空間を学習し，その部分空間の測地線距離を用いる方法です．基本的なアプローチについてはこのMIR2006の論文に書かれています．
SHREC2007の構成とその結果について，詳しくは，
http://www.aimatshape.net/event/SHREC/UU-CS-2007-015.pdf
を参照してください．
Mean Average Precision(relevant) Rank RunFile Value 1 U. of Yamanashi Run 6 0.43371364 2 U. of Yamanashi Run 5 0.43185356 3 ENST-TSI Run 1 0.3787826 4 ITI Run 4 0.33206695 5 ITI Run 3 0.32480213
Our paper that comapred the dimension reduction methods have been presented at the AMR 2007 workshop in Paris, France in July.
Ryutarou Ohbuchi, Jun Kobayashi, Akihiro Yamamoto, and Toshiya Shimizu, Comparison of dimension reduction method for database-adaptive 3D model retrieval, In Proc. Fifth International Workshop on Adaptive Multimedia Retrieval (AMR 2007), Paris, France, July 2007.In Proceedings of the Fifth International Workshop on Adaptive Multimedia Retrieval (AMR 2007), Paris, France, July 2007.(PDF)
Our friend Tatsuma Atsushi has several powerful 3D model retrieval methodls.
It appears that one of his methods would have tied No.1 in the SHREC 2007 Face retrieval contest if his method were to enter the track.
The method achieved NDCG@25 of 0.6379944 using the SHREC 2006 benchmark. This is quite remarkable.
He also has a nifty online 3D model retrieval demo "3Doodle".
My Background
I earned my Ph.D degree at Computer Science Department at the UNC-Chapel Hill where visionaries in computer science teach, and have graduated. Those who teach there include Prof. Frederick. P. Brooks, Prof. Stephen Pizer, Prof. Henry Fuchs, Prof. Turner Whitted, and Prof. Dinesh Manocha. My principal advisor, Prof. Henry Fuchs, is a wonderfully stimulating person to be with. Those who graduated frequent authors lists at the prestigious SIGGRAPH. Chapel Hill is where I switched form computer architecture to computer graphics and worked on volume visualization and augmented reality for medical applications. There is something new andinteresting happening at UNC-Chapel Hill in terms of computer graphics. So please check it out.
From January 1994 to March 1999, I worked for IBM Tokyo Research Laboratory. There, I worked in Advanced Graphics and CAD group. I enjoyed working there, with Masaki Aono, Hiroshi Masuda, Kazunori Miyata, Takayuki Ito, Kenji Shimada, Atushi Yamada, and others. I worked on virtual environment surgical simulation system for National Cancer Center in Tsukiji, Tokyo, realistic image synthesis by using quasi-random sequence, digital watermarking of 3D models, compression techniques for CAD models, and others. .
Contact
Ryutarou Ohbuchi （Japanese -> 大渕 竜太郎）
Computer Science Department
University of Yamanashi
4-3-11 Takeda, Kofu-shi,
Yamanashi-ken, 400-8511
Japan
Phone: +81-55-220-8570
my_last_name AT yamanashi DOT ac DOT jp
