Learning Methods for Generic Object Recognition with Invariance

Learning Methods for Generic Object Recognition with Invariance to Pose and Lighting by Yann Le. Cun, Fu Jie Huang, and Léon Bottou in Proceedings of CVPR'04, 2004 Presentation by Hasan Doğu TAŞKIRAN CS 550 – Machine Learning Department of Computer Engineering Bilkent University April 21, 2005

Outline n n n About the paper… Recognition of Generic Object Categories The NORB Dataset Experiments and Results ¡ Principal Component Analysis ¡ K-Nearest Neighbors ¡ Pairwise Support Vector Machines ¡ Convolutional Networks Conclusion and Future Work 2

The paper is about… n Describing the largest publicly available dataset n Reporting baseline performance with standard methods on this dataset n Exploring how methods fare when the number of input variables is huge n The performance of methods based on global template matching n The performance when the size of the problem is at the upper-limit of applicability n Learning invariance to 3 D pose, lighting conditions and variabilities of images n Taking advantage of binocular inputs 3

Recognition of Generic Object Categories The recognition of generic object categories with invariance to pose, lighting, diverse backgrounds, and the presence of clutter is one of the major challenges of Computer Vision. n n Variety of clues have been used previously: ¡ Color and Texture ¡ Distinctive Local Features ¡ Separately acquired 3 D models ¡ Silhouettes and edges ¡ Pose-invariant Feature Histograms n Shape information? ? 4

Using Shape Information Recognizing Generic Categories such as cars, trucks, airplanes, human figures, or fourlegged animals purely from the shape information is a difficult problem n Another difficulty of the problem is the non-availability of a dataset with sufficient size and diversity to carry out meaningful experiments. n 5

The NORB Dataset n The only useful and reliable clue in the dataset is the shape of the object n NORB is considerably larger than the past datasets and it offers: ¡ More variability ¡ Stereo pairs ¡ The ability to composite the objects and their cast shadows onto diverse backgrounds Images of 50 toys were collected using the peripherals whose details are given in the paper n 6

The NORB Dataset n The collection consists of 10 instance of 5 generic categories: ¡ ¡ ¡ Four-legged Animals, Human Figures, Airplanes, Trucks, Cars All objects are painted uniform green to eliminate irrelevant color and texture Each object instance was placed in a different initial pose 1944 stereo pairs were collected for each instance: 9 elevations, 36 azimuths and 6 lighting conditions A total of 194. 400 images RGB images of resolution 640 x 480 were collected (5 categories, 10 instances, 9 elevations, 36 azimuths, 6 lightings, and 2 cameras) 7

The NORB Dataset Experiments were conducted with 4 datasets generated from the normalized object images n ¡ ¡ Normalized-Uniform Set Jittered-Textured Set Jittered-Cluttered Set Each dataset consists of the 5 instances of categories for training and 5 instances for testing n 8

The NORB Dataset 9

Experiments n On raw image pairs ¡ ¡ n On PCA coefficients ¡ ¡ n Linear Classifier K-Nearest Neighbor Pairwise Support Vector Machines with Gaussian Kernels Convolutional Networks K-Nearest Neighbor Pairwise Support Vector Machines with Gaussian Kernels Lush environment, Torch Library are used 10

Experiments - PCA n 18, 432 x 18, 432 covariance matrix so we need a method Find the principal direction of a centered cloud of points by finding two cluster centroids that are symmetric with respect to the origin i. e. , find u that minimizes n n Yields the first 100 principal components in a few CPU hours 11

Experiments – K-Nearest Neighbors n Running on 24, 300 reference images of size 18, 432 is prohibitively expensive Pre-compute the distances of a few representative images Ak to all other reference images Xi. n Distances are bounded below by: This can be used to choose which distances should be computed first. n Conducted up to K = 18 but best results are obtained for K = 1 12

Experiments – Pairwise SVM Failed to obtain convergence on normalized-uniform dataset in manageable time, also SVMs were not trained on jitter datasets n n Applied on sub-sampled versions and PCA-derived versions 10 SVMs were independently trained to do pairwise classification and used voting strategy n The number of support vectors was between [800, 2000] for PCA-derived inputs n The number of support vectors was between [2000, 3000] for 32 x 32 raw images n 13

Experiments – Convolutional Network Succession of layers of trainable convolutions and spatial sub-sampling n n Extracts features of: ¡ ¡ ¡ Increasingly large receptive fields Increasing complexity Increasing robustness to irrelevant variabilities The network has 90, 575 trainable parameters (Full propagation requires 3, 896, 920 multiply-adds) n Levenberg-Marquardt algorithm with diagonal approximation of the Hessian for 250, 000 online updates n n No over-training, no early-stops. 14

Results 15

Results 16

Results 17

Discussion n These are the first systematic experiments that apply machine learning to shape-based generic object recognition with invariance to pose and lighting n Normalized-uniform dataset is unrealistically favorable to templatebased methods because of the perfect conditions n The size of the jittered database was too large to carry out experiments with the template based methods n The shear size and complexity of the jittered datasets place them above the practical limits of template based methods. n Binocular convolution network take advantage of disparity information to locate the outline of the object 18

Conclusions n n n The system can spot and recognize the animals, human figures, planes, cars and trucks in natural scenes with high accuracy at a rate of several frames per second By presenting the input image at multiple scales, the system can detect those objects over a wide range of scales Popular template-based approaches including SVMs are limited for classification over very large datasets with complex variabilities. Convolutional Networks can be scanned over large images very efficiently The NORB Dataset opens the door to large-scale experiments with learning-based approaches to invariant object recognition Future works may use trainable classifiers that incorporate explicit models of image formation and geometry 19

Comments n The authors just dealt with their problems, not to the specific problems of the algorithms n The paper is well organized and clearly understandable n The dataset preparation details might be reduced n Previous works in the area could be discussed more with their disadvantages 20

Questions? 21