| - Tel:+86-27-87770662
- Fax:+86-27-87770659
- Maggie Shu (data service)
- Benjamin Yao
(tech. support)
|
|
|
If you did not receive our email reply in two weeks after your registration, please
register again! We apologize for the inconvenience! |
 What is the purpose of the LHI dataset? |
| What is inside the LHI dataset ? |
| |
Subset 1 LHI_Transportation_9 |
| |
Subset 2 LHI_UIUC_Sport_Activity_10 |
| |
Subset 3 LHI_UCLA_Aerial_Image_5 |
| |
Subset 4 LHI_Manmade_Object_75 |
| |
Subset 5 LHI_Nature_Object_40 |
| |
Subset 6 LHI_ObjectsinScene |
| |
Subset 7 LHI_SceneSegmentation_18 |
| |
Subset 8 LHI_SurveillanceScenes |
| |
Subset 9 LHI_ImageCategory8 |
| |
Subset 10 LHI_InteractiveSegmentation |
| |
LHI_SceneParsing15Classes |
| How can I download the free dataset? |
| How to use the dataset? |
| Citation and Acknowledgements |
|
Our goal -- To create challenging image databases and benchmarks with high quality human annotated groundtruth.
Image databases are an essential element of Computer Vision research. Current datasets, such as the 'Caltech 101', however, offer a somewhat limited range of image variability. As commented by James DiCarlo from MIT (link), "These image sets have design flaws that enable computers to succeed where they would fail with more authentically varied images". Another paper named "Dataset issue in object recognition" by J. Ponce et. al, also put current datasets under inspection. Indeed, the "dataset issue" is a big challenge against every researcher who takes Computer Vision seriously. There are dozens of problems remain unanswered, such as: How to build a general image database without bias to purpose? How to create benchmark that reflects the real-world difficulty of image understanding? How to guarantee the correctness of annotation? etc. The LHI dataset project is started with the surgence of need for better dataset and now, after 2 years, collaboration between a team of 23 full-time labelers at the Lotus Hill Research
Institute (LHI), China, and the UCLA Center for Image and Vision
Sciences (CIVS) , we have built up the LHI dataset with annotations covering quite broad scope and high accuracy. Although we still don't have answers to all those questions above, we are stepping toward these directions and doing experiments along the way. |
|
|
|
| A quick preview to the LHI dataset
The table below gives you an overview of the subsets and what kind of annotation information comes together with each subset. From left, the first column displays the names of all 8 subsets. The 2nd to 8th columns correspond to seven annotation methods (click on the name will lead you to a brief description of the method, more detailed information can be found from our paper). Letter "Y" means a subset has certain type of annotation, "N" means has not. (For example, subset 1 "LHI_Transportation_10" has segmentation, sketch map, hierarchical decomposition, templates and are labeled under multiple views&scales, but does not have ground planeand/or horizon line) |
| Subsets |
Segmentation |
SketchMap |
|
Templates |
GroundPlane and/or Horizon line |
Multiple Views |
Multiple Scales |
|
Y |
Y |
Y |
Y |
N |
Y |
Y |
|
Y |
Y |
Y |
Y |
Y |
N |
N |
|
Y |
N |
N |
N |
N |
N |
N |
|
Y |
Y |
Y |
N |
N |
Y |
Y |
|
Y |
Y |
Y |
N |
N |
Y |
Y |
|
Y |
Y |
Y |
Y |
N |
N |
N |
|
Y |
N |
N |
N |
N |
N |
N |
|
|
|
|
|
|
|
|
Table.1 |
|
|
|
|
|
|
|
|
|
|
How can I download the free dataset?
The dataset and Matlab toolbox is free to researchers at academic institutions for non-commercial purpose. Please click download to enter a registration and downloading page. |
|
View, Download and Customize the Dataset with the Matlab ToolBox
Using the provided Matlab ToolBox, you can view the dataset online or download to your local machine. Click here for get a detailed instruction on how to use the Matlab Toolbox and download the datasets.
|
|
|
|
| Citation
If you use this dataset, please cite the paper: Benjamin Yao, Xiong Yang, and Song-Chun Zhu, "Introduction to a large scale general purpose ground truth dataset: methodology, annotation tool, and benchmarks." EMMCVPR, 2007(pdf).
We would like to thank to following |
Acknowledgement
We would like to thank to the LabelMe dataset for providing valuable experience and code on constructing and organizing dataset.
We also would like to thank to Alan Yuille, Li Fei-Fei, Antonio Tarrabal, David Martin, Charles Fowlkes, Peter Hallinan, Mark Nitzberg for provide their valuable suggestions at the Kick-off meeting in LHI 2005. Over the years, the following vision experts have taught summer courses at Lotus Hill: Long Quan, Jackie Shen, Ying Wu, Zhuowen Tu, Yi Ma, Jianbo Shi, Stan Li, Changshui Zhang. |
|
- Lotus Hill Reseach Institute for Computer Vision and
Information Science, City of Wuhan, Hubei 430074, China
- 1024*768 IE6.0
- Copyright 2007 by
www.imageparsing.com. All rights Reserved.
|