CarData.zip

--------------------------------------------------- UIUC Image Database for Car Detection : README File --------------------------------------------------- Besides this README file, the CarData/ directory (to which the download .tar file unpacks) contains the following: 1. Training images ------------------ These are located in CarData/TrainImages/. The positive (car) images are named pos-n.pgm (0 <= n <= 549). The negative (non-car) images are named similarly as neg-n.pgm (0 <= n <= 499). The training images are all 100x40 pixels in size. 2. Test images - Single-scale ----------------------------- These are located in CarData/TestImages/ and are named test-n.pgm (0 <= n <= 169). The images are of different sizes themselves but contain cars of approximately the same scale as in the training images. These test images contain 200 cars in all. 3. Test images - Multi-scale ---------------------------- These are located in CarData/TestImages_Scale/ and are named test-n.pgm (0 <= n <= 107). The images are of different sizes and contain cars of various scales. These test images contain 139 cars in all. 4. Evaluation files - Single-scale ---------------------------------- There are four files: (a) trueLocations.txt This contains the true locations of the objects (cars) in the first set of test images. The locations have been determined manually and are represented by the coordinates of the top-left corner of the best 100x40 window containing each car. The nth row in the file corresponds to the nth test image; for example, if the nth test image contains two cars located at (i1,j1) and (i2,j2) respectively (where i denotes the height axis and j the width axis in usual image matrix notation), then the nth row looks like: n: (i1,j1) (i2,j2) (b) Point.class This file is simply required by the evaluator program; it can be ignored. (c) Evaluator.java This contains the source code for the evaluator program; it is provided to the user in case some modification is desired, but if using the evaluation scheme as provided, this file can be ignored. (d) Evaluator.class This can be used to evaluate the performance of a detection algorithm on the test set. If the object locations determined by the algorithm are output to a file "foundLocations.txt" (which must be in the same format as the "trueLocations.txt" file described in part (a) above), the evaluation can be performed by running the command ~CarData> java Evaluator trueLocations.txt foundLocations.txt This determines the number of correct and false detections, and outputs the results in the form of recall, precision and F-measure. It also produces a file called "break-up.txt" which contains a break-up of the correct and false detections for each test image. (A location output by the algorithm is counted as a correct detection if it lies within an ellipse with center at the true location and axes 25% of the object dimensions in each direction. In addition, only one detection per object is allowed - if two or more detected windows satisfy the above criteria for the same object, only one is counted as correct; the others are counted as false detections.) The evaluator program can also be used to display the locations output by the algorithm as windows overlaid on the test images; correct detections are shown as complete windows, false detections as broken windows. This can be done by creating a directory, say "ResultImages", under CarData/ and then running: ~CarData> java Evaluator trueLocations.txt foundLocations.txt display TestImages ResultImages The results are produced as images named "result-n.pgm" that can be found in CarData/ResultImages/. This feature can also be used to display the test images with the true object locations; create a directory "LabeledImages" under CarData/ and run: ~CarData> java Evaluator trueLocations.txt trueLocations.txt display TestImages LabeledImages The labeled images are produced as CarData/LabeledImages/result-n.pgm. 5. Evaluation files - Multi-scale --------------------------------- There are four files: (a) trueLocations_Scale.txt This contains the true locations and scales of the objects (cars) in the second set of test images. The locations and scales have been determined manually and are represented by the coordinates of the top-left corner and the width of the best window containing each car. The nth row in the file corresponds to the nth test image; for example, if the nth test image contains two cars with location-scale pairs (i1,j1,w1) and (i2,j2,w2) respectively (where i denotes the height axis and j the width axis in usual image matrix notation), then the nth row looks like: n: (i1,j1,w1) (i2,j2,w2) (b) Point_Scale.class This file is simply required by the evaluator program; it can be ignored. (c) Evaluator_Scale.java This contains the source code for the evaluator program; it is provided to the user in case some modification is desired, but if using the evaluation scheme as provided, this file can be ignored. (d) Evaluator_Scale.class This can be used to evaluate the performance of a detection algorithm on the test set. If the object locations and scales determined by the algorithm are output to a file "foundLocations_Scale.txt" (which must be in the same format as the "trueLocations_Scale.txt" file described in part (a) above), the evaluation can be performed by running the command ~CarData> java Evaluator_Scale trueLocations_Scale.txt foundLocations_Scale.txt This determines the number of correct and false detections, and outputs the results in the form of recall, precision and F-measure. It also produces a file called "break-up_Scale.txt" which contains a break-up of the correct and false detections for each test image. (A location-scale pair output by the algorithm is counted as a correct detection if it lies within an ellipsoid with center at the true location-scale and axes 25% of the true object dimensions in each direction. In addition, only one detection per object is allowed - if two or more detected windows satisfy the above criteria for the same object, only one is counted as correct; the others are counted as false detections.) The evaluator program can also be used to display the locations output by the algorithm as windows overlaid on the test images; correct detections are shown as complete windows, false detections as broken windows. This can be done as described for the single-scale case in 4(d) above. -------------------------------------------------------------------------------