IEEE Conference on Computer Vision and Pattern Recognition, San Juan, Puerto Rico, June 1997 Object Detection with Vector Quantized Binary Features John Krumm Intelligent Systems & Robotics Center Sandia National Laboratories Albuquerque, NM 87185 jckrumm@sandia.gov Abstract This paper presents a new algorithm for detecting objects in images, one of the fundamental tasks of computer vision. The algorithm extends the representational efficiency of eigenimage methods to binary features, which are less sen- sitive to illumination changes than gray-level values nor- mally used with eigenimages. Binary features (square subtemplates) are automatically chosen on each training image. Using features rather than whole templates makes the algorithm more robust to background clutter and par- tial occlusions. Instead of representing the features with real-valued eigenvector principle components, we use bi- nary vector quantization to avoid floating point computa- tions. The object is detected in the image using a simple geometric hash table and Hough transform. On a test of 1000 images, the algorithm works on 99.3%. We present a theoretical analysis of the algorithm in terms of the re- ceiver operating characteristic, which consists of the prob- abilities of detection and false alarm. We verify this analy- sis with the results of our 1000-image test, and we use the analysis as a principled way to select some of the algo- rithm’s important operating parameters. 1. Overview and Context Detecting objects in images and measuring their location is a fundamental task of computer vision, with applications in manufacturing, inspection, world modeling, and target recognition. Often the scene is inherently cluttered, the object may be partially occluded, and illumination may change. In this case, the algorithm must look at features internal to the objects’ silhouette, and look at them in such a way that missing features and changing illumination are tolerated. Researchers have responded to this need in many ways, including fairly recent, elegant object detection algorithms based on principle components of training images of the object[[6]][[10]]. In particular, Murase and Nayar[[6]] extract templates from training images of the object in dif- ferent orientations, compute eigenvector principle compo- nents of these templates, and recover the object’s orienta- tion in new images by projecting them onto the principle components. They address the problem of illumination changes by taking training images under different lighting This work was performed at Sandia National Laboratories and supported by the U.S. Department of Energy under contract DE-AC04-94AL85000. conditions. The whole-object template idea was improved by algorithms that look at only part[[7]] or parts[[4]][[8]] of the training templates. This helped to reduce or elimi- nate the effects of background clutter and partial occlu- sions. This paper presents and analyzes a new algorithm for object detection based on binary subtemplates (features) of the training images. Binary features are more robust to illumination changes than the gray-level features of previ- ous methods. We replace the eigenvector principle com- ponents with binary vector quantization, a common method for image compression. This avoids any floating point processing after edge detection. An object is represented with a separate, distinct model for each pose in the training set, which avoids problems caused by self-occlusion and the movement of specular highlights. A model consists of vector quantized binary features and their relative spatial offsets with respect to each other. An example of our de- tection results in an image with background clutter and partial occlusions is shown in Figure 1. On a test of 1000 cluttered images containing the test object, the algorithm correctly detected the object in 99.3%. In order to deter- mine the best values of the algorithm’s important parame- ters, we derive and plot the receiver operating characteris- Figure 1: Result of detection algorithm in presence of background clutter and partial occlusions.