Caltech 101

From Wikipedia, the free encyclopedia

Caltech 101 is a dataset of digital images created in September, 2003, compiled by Fei-Fei Li, Marco Andreetto, and Marc 'Aurelio Ranzato at the California Institute of Technology. It is intended to facilitate Computer Vision research and techniques. It is most applicable to techniques interested in recognition, classification, and categorization. Caltech 101 contains a total of 9146 images, split between 101 distinct object (including faces, watches, ants, pianos, etc.) and a background category (for a total of 102 categories). Provided with the images are a set of annotations describing the outlines of each image, along with a Matlab script for viewing.

Contents

[edit] Purpose

Most Computer Vision and Machine Learning algorithms function by training on a large set of example inputs. To work effectively, most of these techniques require a large and varied set of training data. For example, the relatively well known real time face detection method used by Paul Viola and Micheal J. Jones was trained on 4916 hand labeled faces [1]. However, acquiring a large volume of appropriate and usable images is often difficult. Furthermore, cropping and resizing many images, as well as marking point of interest by hand, is a tedious and time intensive task.

Historically, most datasets used in computer vision research have been tailored to the specific needs of the project being worked on.

Image:Caltech101vs256.gif
Caltech 101 vs Caltech 256 on same algorithms

A large problem in comparing different computer vision techniques is the fact that most groups are using their own datasets. Each of these datasets may have different properties that make reported results from different methods harder to compare directly. For example, differences in image size, image quality, relative location of objects within the images, and level of occlusion and clutter present can lead to varying results.

The Caltech 101 dataset aims to alleviate many of these common problems.

  • The work of collecting a large set of images, and cropping and resizing them appropriately has been taken care of.
  • A large number of different categories are represented, which benefits both single, and multi class recognition algorithms.
  • Detailed object outlines have been marked for each image.
  • By being released for general use, the Caltech 101 acts as a common standard by which to compare different algorithms without bias due to different datasets.

However, a recent study [2] demonstrates that tests based on uncontrolled natural images (like the Caltech 101 dataset) can be seriously misleading, potentially guiding progress in the wrong direction.

[edit] The Dataset

[edit] Images

Image:Caltech101.gif
Caltech 101 images

The Caltech 101 dataset consists of a total of 9146 images, split between 101 different object categories, as well as an additional background/clutter category.

Each object category contains between 40 and 800 images on average. Common and popular categories such as faces tend to have a larger number of images than less used categories. Each image is about 300x200 pixels in dimension. Images of oriented objects such as airplanes and motorcycles were mirrored to be left-right aligned, and vertically oriented structures such as buildings were rotated to be off axis.

[edit] Annotations

As a supplement to the images, a set of annotations are provided for each image. Each set of annotations contains two pieces of information.

The general bounding box in which the object is located, and a detailed human specified outline enclosing the object. A Matlab script is provided along with the annotations that will load an image and its corresponding annotation file and display them as a Matlab figure.

Image:Caltech101 croc annotated.jpg

The bounding box is yellow and the outline is red.

[edit] Uses

The Caltech 101 dataset has been used to train and test several Computer Vision recognition and classification algorithms. The first paper to make use of Caltech 101 was an incremental Bayesian approach to one shot learning [3]. One shot learning is an attempt to learn a class of object using only a few examples, by building off of prior knowledge of many other classes.

The Caltech 101 images, along with the annotations, were used for another one shot learning paper at Caltech.

L. Fei-Fei, R. Fergus and P. Perona. One-Shot learning of object categories [4]

Other Computer Vision papers that report using the Caltech 101 dataset:

  • Shape Matching and Object Recognition using Low Distortion Correspondence. Alexander C. Berg, Tamara L. Berg, Jitendra Malik. CVPR 2005
  • The Pyramid Match Kernel:Discriminative Classification with Sets of Image Features. K. Grauman and T. Darrell. International Conference on Computer Vision (ICCV), 2005 [5]
  • Combining Generative Models and Fisher Kernels for Object Class Recognition Holub, AD. Welling, M. Perona, P. International Conference on Computer Vision (ICCV), 2005 [6]
  • Object Recognition with Features Inspired by Visual Cortex. T. Serre, L. Wolf and T. Poggio. Proceedings of 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005), IEEE Computer Society Press, San Diego, June 2005.[7]
  • SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition. Hao Zhang, Alex Berg, Michael Maire, Jitendra Malik. CVPR, 2006[8]
  • Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories. Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. CVPR, 2006[9]
  • Empirical study of multi-scale filter banks for object categorization, M.J. Mar韓-Jim閚ez, and N. P閞ez de la Blanca. December 2005[10]
  • Multiclass Object Recognition with Sparse, Localized Features, Jim Mutch and David G. Lowe. , pg. 11-18, CVPR 2006, IEEE Computer Society Press, New York, June 2006[11]
  • Using Dependent Regions or Object Categorization in a Generative Framework, G. Wang, Y. Zhang, and L. Fei-Fei. IEEE Comp. Vis. Patt. Recog. 2006[12]

[edit] Analysis and Comparison

[edit] Advantages

Caltech 101 has several advantages over other similar datasets:

  • Uniform size and presentation.

Almost all the images within each category are uniform in image size and in the relative position of interest objects. This means that, in general, users who wish to use the Caltech 101 dataset do not need to spend and extra time cropping and scaling the images before they can be used.

  • Low level of clutter/occlusion:

Algorithms concerned with recognition usually function by storing features unique to the object that is to be recognized. However, the majority of images taken have varying degrees of background clutter. Algorithms trained on cluttered images can potentially build incorrect

  • Detailed Annotations:

The detailed annotations of object outlines is another advantage to using the dataset.

[edit] Weaknesses

There are several weaknesses to the Caltech 101 dataset [2] [13]. Some of them are conscious trade-offs for the advantages it provides, and some are simply limitations of the dataset itself.

  • Limited number of categories:

There are approximately 10,000 different categories of objects. The Caltech 101 dataset represents only a small fraction of these.

  • Some categories contain few images:

Certain categories are not represented as well as others, containing as few as 31 images. This means that \mathrm{N}_{\mathrm{train}} \le 30. The number of images used for training must be less than or equal to 30, which is not sufficient for all purposes.

  • Can be too easy:

Images are very uniform in presentation, left right aligned, and usually not occluded. As a result, the images are not always representative of practical inputs that the algorithm being trained might be expected to see. Under practical conditions, there is usually more clutter, occlusion, and variance in relative position and orientation of interest objects.

  • Aliasing and artifacts due to manipulation:

Some images have been rotated and scaled from their original orientation, and suffer from some amount of artifacts or aliasing.

[edit] Other Datasets

  • Caltech 256 is another image dataset created at the California Institute of technology in 2007, a successor to Caltech 101. It is intended to address some of the weaknesses inherent to Caltech 101. Overall, it is a more difficult dataset than Caltech 101 (but it suffers from the same problems [2])
    • 30,607 images, covering a larger number of categories.
    • Minimum number of image per category raised to 80.
    • Images not left-right aligned.
    • More variation in image presentation.
  • LabelMe is an open, dynamic dataset created at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). LabelMe takes a different approach to the problem of creating a large image dataset, with different trade-offs.
    • 106,739 images, 41,724 annotated images, and 203,363 labeled objects.
    • Users may add images to the dataset by upload, and add labels or annotations to existing images.
    • Due to its open nature, LabelMe has many more images covering a much wider scope than Caltech 101. However, since each person decides what images to upload, and how to label and annotate each image, there can be a lack of consistency between images.

[edit] References

  1. ^ P. Viola and M. J. Jones, Robust Real-Time Object Detection, , IJCV 2004
  2. ^ a b c | Why is Real-World Visual Object Recognition Hard? Pinto N, Cox DD, DiCarlo JJ PLoS Computational Biology Vol. 4, No. 1, e27 doi:10.1371/journal.pcbi.0040027
  3. ^ |L. Fei-Fei, R. Fergus and P. Perona. Learning generative visual models from few training examples: an incremental Bayesian approach tested on 101 object categories. IEEE. CVPR 2004, Workshop on Generative-Model Based Vision. 2004
  4. ^ | L. Fei-Fei, R. Fergus and P. Perona. One-Shot learning of object categories. IEEE Trans. Pattern Analysis and Machine Intelligence, Vol28(4), 594 - 611, 2006.
  5. ^ | The Pyramid Match Kernel:Discriminative Classification with Sets of Image Features. K. Grauman and T. Darrell. International Conference on Computer Vision (ICCV), 2005
  6. ^ | Combining Generative Models and Fisher Kernels for Object Class Recognition Holub, AD. Welling, M. Perona, P. International Conference on Computer Vision (ICCV), 2005
  7. ^ | Object Recognition with Features Inspired by Visual Cortex. T. Serre, L. Wolf and T. Poggio. Proceedings of 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005), IEEE Computer Society Press, San Diego, June 2005
  8. ^ | SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition. Hao Zhang, Alex Berg, Michael Maire, Jitendra Malik. CVPR, 2006
  9. ^ | Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories. Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. CVPR, 2006
  10. ^ | Empirical study of multi-scale filter banks for object categorization, M.J. Mar韓-Jim閚ez, and N. P閞ez de la Blanca. December 2005
  11. ^ | Multiclass Object Recognition with Sparse, Localized Features, Jim Mutch and David G. Lowe. , pg. 11-18, CVPR 2006, IEEE Computer Society Press, New York, June 2006
  12. ^ | Using Dependent Regions or Object Categorization in a Generative Framework, G. Wang, Y. Zhang, and L. Fei-Fei. IEEE Comp. Vis. Patt. Recog. 2006
  13. ^ | Dataset Issues in Object Recognition. J. Ponce, T. L. Berg, M. Everingham, D. A. Forsyth, M. Hebert, S. Lazebnik, M. Marszalek, C. Schmid, B. C. Russell, A. Torralba, C. K. I. Williams, J. Zhang, and A. Zisserman. Toward Category-Level Object Recognition, Springer-Verlag Lecture Notes in Computer Science. J. Ponce, M. Hebert, C. Schmid, and A. Zisserman (eds.), 2006

[edit] External links