Object Indexing Using an Iconic Sparse Distributed Memory

Rajesh P. N. Rao, Dana H. Ballard

Abstract

A general-purpose object indexing technique is described that combines the virtues of principal component analysis with the favorable matching properties of high-dimensional spaces to achieve high-precision recognition. An object is represented by a set of high-dimensional iconic feature vectors comprised of the responses of derivatives of Gaussian filters at a range of orientations and scales. Since these filters can be shown to form the eigenvectors of arbitrary images containing both natural and man-made structures, they are well-suited for indexing in disparate domains. The indexing algorithm uses an active vision system in conjunction with a modified form of Kanerva's (1988, 1993) sparse distributed memory which facilitates interpolation between views and provides a convenient platform for learning the association between an object's appearance and its identity. The robustness of the indexing method was experimentally confirmed by subjecting the method to a range of viewing conditions and the accuracy was verified using a well-known model database containing a number of complex 3D objects under varying pose.

Introduction

Traditional computer vision has often relied on geometric primitives—edges, vertices, and 3D wireframes—to recognize objects. However, extracting these features from real-world images is frequently hampered by complex lighting, shadows, and textures. Rao and Ballard propose an alternative "iconic" approach: representing objects by their local photometric appearance. The core challenge of this approach lies in indexing: efficiently storing and retrieving high-dimensional appearance vectors in a way that is robust to environmental changes.

Multiscale Iconic Features

The system extracts features using a bank of nine derivative-of-Gaussian filters (up to the 3rd order) across five scales. These filters are not arbitrary; the authors demonstrate that they emerge naturally when performing Principal Component Analysis (PCA) on natural images, suggesting they are an optimal basis for visual representation. To maintain invariance to rotation, the system employs "steerable filters," which allow the feature vector to be mathematically rotated to a canonical orientation without re-scanning the image.

Multiscale derivative-of-Gaussian filters used for feature extraction across different scales.

Modified Sparse Distributed Memory (SDM)

To manage the resulting high-dimensional vectors (typically 360+ dimensions), the researchers adapted Pentti Kanerva’s Sparse Distributed Memory. While the original SDM was designed for binary data, this modified version utilizes continuous-valued vectors and employs normalized dot products (correlation) as a distance metric.

This architecture treats associative memory as a "memory surface." By leveraging the mathematical properties of high-dimensional spaces—where most vectors are naturally nearly orthogonal—the system ensures that objects can undergo significant distortion or noise before they are confused with other stored memories. This search process is implemented as a series of convolutions, enabling real-time performance at 30 fps on dedicated hardware.

The architecture of the modified Sparse Distributed Memory (SDM) showing the indexing and retrieval process.

Experimental Results and Robustness

The system was evaluated using the Columbia Object Database, which contains 20 complex 3D objects. Key findings include:

  • High Precision: The system achieved a 100% recognition rate when using 25 iconic points per object. Even with a single point, accuracy remained at 70%.
  • Pose Invariance: Iconic vectors remained highly correlated (>0.8) even when the 3D pose of an object changed by as much as 40°.
  • Graceful Degradation: In "worst-case" tests involving simultaneous occlusion, illumination changes, and scale variations, the correlation remained above 0.5—well above the noise floor of unrelated objects (near 0.0).

Conclusion

By combining biologically-inspired iconic features with the robust matching properties of high-dimensional associative memory, this research provides a powerful framework for general-purpose object recognition. Future work aims to improve scalability to thousands of objects and implement saliency-based point selection to automatically identify the most unique features of an object for indexing.

Summary updated

Abstract

A general-purpose object indexing technique is described that combines the virtues of principal component analysis with the favorable matching properties of high-dimensional spaces to achieve high-precision recognition. An object is represented by a set of high-dimensional iconic feature vectors comprised of the responses of derivatives of Gaussian filters at a range of orientations and scales. Since these filters can be shown to form the eigenvectors of arbitrary images containing both natural and man-made structures, they are well-suited for indexing in disparate domains. The indexing algorithm uses an active vision system in conjunction with a modified form of Kanerva's (1988, 1993) sparse distributed memory which facilitates interpolation between views and provides a convenient platform for learning the association between an object's appearance and its identity. The robustness of the indexing method was experimentally confirmed by subjecting the method to a range of viewing conditions and the accuracy was verified using a well-known model database containing a number of complex 3D objects under varying pose.<
BibTeX
@inproceedings{rao1995object,
  author = {Rajesh P. N. Rao and Dana H. Ballard},
  title = {{Object Indexing Using an Iconic Sparse Distributed Memory}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {1995},
  url = {https://openreview.net/forum?id=r1bGx-fuZr},
}
Figure