Histograms of Oriented Gradients for Human Detection(HOG) Dalal, N.; Triggs, B., IEEE Computer Society Conference on Computer Vision and Pattern Recognition(2005) vol. 1 ,pp.886 - 893 Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen
Outline 1. Introduction 2. Methodology 3. Results 4. Discussion 5. Conclusion
Introduction Detecting humans in images is a challenging task Variable appearance Wide range of poses A robust feature set Discriminate cleanly Cluttered backgrounds Different illumination
Introduction Edge orientation histograms Scale-invariant feature transform (SIFT) Shape context SIFT Shape context
Introduction Using linear SVM as a baseline classifier Using detection error tradeoff (DET) Data Sets MIT pedestrian set INRIA pedestrian set
Methodology Data Sets MIT pedestrian database INRIA 509 training images 200 test images INRIA 1805 64X128 images
Methodology 1 2 3 4 5 6 7 8 9 10
Methodology Training examples 12180+ examples 2478 Positive 1218 Negative
Methodology Detection error tradeoff X-axes Y-axes Log-log scale False Positives Per Window tested(by 5% at 10-4) FPPW= Y-axes Miss rate= Log-log scale
Methodology
Gamma/Color Normalization Inputting pixel representations Grayscale RGB color spaces LAB color spaces Power law (Gamma equalization)
LAB Color Spaces Xn, Yn and Zn are the CIE XYZ tristimulus values of the reference white point
Power Law (Gamma equalization) Tradition IGray(i, j) is the gray-level image IEq (i, j) is the image which performed equalization IMax and IMin are the maximum and minimum of the pixel values of IGray(i, j)
Power Law (Gamma equalization) i is the i-th gray level L is the low-bound R is the actual equalization range GE (i) is the result of the i-th gray level obtained from gamma equalization
Power Law (Gamma equalization)
Gradient Computation Masks test(for each color channel) Gaussian (σ=0~3) 1-D point derivatives[-1,0,1] Cubic-corrected[1,-8,0,8,-1] 3X3 Sobel mask 2X2 diagonal ones
Gradient Computation ‘c-cor’ is the 1D cubic-corrected point derivative
Spatial / Orientation Binning Orientation bins are evenly spaced 0 °~180 ° 0 °~360 °
Spatial / Orientation Binning
Normalization and Descriptor Blocks
Normalization and Descriptor Blocks
Normalization and Descriptor Blocks Block Normalization schemes (limiting the maximum values of v to 0.2) and renormalizing Centre-surround normalization Window norm(using Gaussian σ=1) v is the unnormalized descriptor vector is a small constant
Normalization and Descriptor Blocks
Normalization and Descriptor Blocks Illumination and foreground-background contrast overlap
Normalization and Descriptor Blocks
Detector Window and Context
Classifier Using linear SVM(Support vector machine) Increasing performance Using a Gaussian kernel Higher run time
Classifier Using a Gaussian kernel SVM,
Results
Results
Results The performance of selected detectors on (left) MIT and (right) INRIA data sets.
Discussion HOG outperform wavelet & shape context Traditional centre-surround style schemes are not the best choice Similar to SIFT descriptors
Conclusion Scale gradients Orientation binning Relatively coarse spatial binning High-quality local contrast normalization in overlapping descriptor blocks
Thank you for your attention