120 likes | 271 Views
Object Recognition. Article: Distinctive Image Features from Scale -Invariant Keypoints. C. Pantofaru and M. Hebert. CMU Technical Report, September 2005. Xiang Li
E N D
Object Recognition Article: Distinctive Image Features from Scale -Invariant Keypoints. C. Pantofaru and M. Hebert. CMU Technical Report, September 2005. Xiang Li Jan, 21, 2009
Introduction • Scale-space peak selection • Peak detection • Frequency of sampling in scale • Frequency of sampling in spatial domain • Keypoint localization • Orientation assignment • Keypoint descriptor • Application
Scale-space peak selection • difference-of-Gaussian function D(x, y, σ) = (G(x, y, k) - G(x, y, σ)) * I(x; y) = L(x, y, k) - L(x, y, σ) Q: Why there is a need to difference Gaussian function instead of pure Gaussian function? (Mahdi)
Peak detection • To detect the locations of all local maxima and minima (peaks) of D(x, y, σ) • Using the sampling frequency which provide the best results under a realistic simulation.
Frequency of sampling in scale • The highest repeatability is obtained when sampling 3 scales per octave.
Frequency of sampling in spatial domain • Repeatability and Cost • we have chosen to use σ = 1.6, which provides close to optimal repeatability. Q : There is a cost to using a large σ in terms of efficiency and a reduced number of keypoints (shown by the lower line). How did he get the conclusion?
Keypoint localization • Taylor expansion • For the experiments in this paper, all peaks with a value of D(^x) less than 0.03 were discarded.
Orientation assignment The experimental stability of orientation assignment under differing amounts of image noise.
Keypoint descriptor • Based on a model of biological vision • Different computational mechanism • Descriptor representation • Descriptor testing • Sensitivity to affine change • Matching to large databases
Application • Keypoint matching • Efficient nearest neighbor indexing Q: What are bins which are to be searched in Best-Bin-First (BBF) algorithm in feature space? (Dustin) • A modification of the k-d tree algorithm called the best-bin-first search method can identify the nearest neighbors with high probability using only a limited amount of computation. Q: In Page 21, it said if the ratio is greater than 0.8, it should be rejected. How did he get the number 0.8? (Xiang LI)
Question Q: What is Hough transform? (Mahdi) • It is used to search for keys that agree upon a particular model pose. • A feature extraction technique used in image analysis, computer vision, and digital image processing. • It is to find imperfect instances of objects within a certain class of shapes by a voting procedure. • This voting procedure is carried out in a parameter space, from which object candidates are obtained as local maxima in a so-called accumulator space that is explicitly constructed by the algorithm for computing the Hough transform. • Ballard, D.H., “Generalizing the Hough transform to detectarbitrary patterns,” Pattern Recognition, 13, 2 (1981), pp.111-122.