【正文】 plate image and should be chosen such that not all N matches are around a single character when the same character occurs more than once on a plate, and not too large so that not all possible regions are processed. This method may seem inefficient, however, the recognition process takes on the order of half a second for a resolution of 104 31, which we found to be acceptable. 4 Datasets We automatically generated a database of car images by running our license plate detector and tracker on several hours of video data and cropping a fixed window of size 400 220 pixels around the license plate of the middle frame of each tracked sequence. This method yielded 1,140 images in which cars of each make and model were of roughly the same size. The crop window was positioned such that the license plate was centered in the bottom third of the image. We chose this position as a reference point to ensure matching was done with only car features and not background features. Had we centered the license plate both vertically and horizontally, cars that have their plates mounted on their bumper would have exposed the road in the image. After collecting these images, we manually assigned make, model, and year labels to 790 of the 1,140 images. We were unable to label the remaining 350 images due to our limited familiarity with those cars. We often made use of the California Department of Motor Vehicles’ web site to determine the makes and models of cars with which we were not familiar. The web site allows users to enter a license plate or vehicle identification number for the purposes of checking whether or not a car has passed recent smog checks. For each query, the web site returns smog history as well as the car’s make and model description if available. The State of California requires all vehicles older than three years to pass a smog check every two years. Therefore, we were unable to query cars that were three years old or newer and relied on our personal experience to label them. We split the 1,140 labeled images into a query set and a database set. The query set contains 38 images chosen to represent a variety of make and model classes, in some cases with multiple queries of the same make and model but different year in order to capture the variation of model designs over time. We evaluated the performance of each of the recognition methods by finding the best match in the database for each of the query images. SIFT Matching Scale invariant feature transform (SIFT) features recently developed by Lowe [14] are invariant to scale, rotation and even partially invariant to illumination differences, which makes them well suited for object recognition. We applied SIFT matching to the problem of MMR as follows: 1. For each image d in the database and a query image q, perform keypoint localization and descriptor assignment. 2. For each database image d: (a) For each keypoint kq in q find the keypoin