Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.

Gradient-based learning applied to document recognition

We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. By a carefully crafted design, we increased the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC14 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

/pdf/going-deeper-with-convolutions-1yobw2o2ds.pdf

Going deeper with convolutions

There is, I think, something ethereal about i —the square root of minus one. I remember first hearing about it at school. It seemed an odd beast at that time—an intruder hovering on the edge of reality.

Usually familiarity dulls this sense of the bizarre, but in the case of i it was the reverse: over the years the sense of its surreal nature intensified. It seemed that it was impossible to write mathematics that described the real world in …

I and i

We study the question of feature sets for robust visual object recognition; adopting linear SVM based human detection as a test case. After reviewing existing edge and gradient based descriptors, we show experimentally that grids of histograms of oriented gradient (HOG) descriptors significantly outperform existing feature sets for human detection. We study the influence of each stage of the computation on performance, concluding that fine-scale gradients, fine orientation binning, relatively coarse spatial binning, and high-quality local contrast normalization in overlapping descriptor blocks are all important for good results. The new approach gives near-perfect separation on the original MIT pedestrian database, so we introduce a more challenging dataset containing over 1800 annotated human images with a large range of pose variations and backgrounds.

/pdf/histograms-of-oriented-gradients-for-human-detection-3nteinziuq.pdf

Histograms of oriented gradients for human detection

Object detection performance, as measured on the canonical PASCAL VOC dataset, has plateaued in the last few years. The best-performing methods are complex ensemble systems that typically combine multiple low-level image features with high-level context. In this paper, we propose a simple and scalable detection algorithm that improves mean average precision (mAP) by more than 30% relative to the previous best result on VOC 2012 -- achieving a mAP of 53.3%. Our approach combines two key insights: (1) one can apply high-capacity convolutional neural networks (CNNs) to bottom-up region proposals in order to localize and segment objects and (2) when labeled training data is scarce, supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost. Since we combine region proposals with CNNs, we call our method R-CNN: Regions with CNN features. We also present experiments that provide insight into what the network learns, revealing a rich hierarchy of image features. Source code for the complete system is available at http://www.cs.berkeley.edu/~rbg/rcnn.

/pdf/rich-feature-hierarchies-for-accurate-object-detection-and-22fvjlggyg.pdf

Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation

One of the fundamental aspects of vision is three-dimensional geometry referring to the relationship between the scene which is depicted and the images which are projections of that scene. This paper presents computational theories which deal with geometrical aspects of interpreting images as a three-dimensional (3D) scene. The presentation covers three levels of descriptions: micro-surface level, volumetric-object level, and scene level. The micro-surface level concerns the relationship between surface orientations of small surface patches and image properties. Theories are presented which provide computational constraints for recovering surface orientations from geometrical image properties, such as symmetry, texture, shadows, and perspective distortions. The volumetric-object level handles shape representations of primitive objects: the relationship between 3D shape of an object and its projections onto images must be understood. The generalized cylinder is a popular volumetric shape representation in vision. We will discuss its formal properties. Finally, the scene level aims at constructing and maintaining the 3D description of the whole scene. We will present the Incremental 3D Mosaic system under development at CMU, which incrementally constructs the total 3D scene description of the task area from aerial photographs.

/pdf/geometrical-aspects-of-interpreting-images-as-a-three-41es1xij4u.pdf

Geometrical aspects of interpreting images as a three-dimensional scene

Computer vision techniques can estimate 3D shape from images, but usually only up to a scale factor. The scale factor must be obtained by a physical measurement of the scene or the camera motion. Using gauge theory, we show that how this scale factor is determined can significantly affect the accuracy of the estimated shape. And yet these considerations have been ignored in previous works where 3D shape accuracy is optimized. We investigate how scale fixing influences the accuracy of 3D reconstruction and determine what measurement should be made to maximize the shape accuracy.

/pdf/gauge-fixing-for-accurate-3d-estimation-343jbdo7dq.pdf

Gauge fixing for accurate 3D estimation

Digitally recording dynamic events, such as sporting events, for experiencing in a spatio-temporally distant and arbitrary setting requires 4D capture: three dimensions for their geometry and appearance over the fourth dimension of time. Today's computer vision techniques make 4D capture possible. The virtualized reality system serves as an example on the general problem of digitizing dynamic events. In this article, we present the virtualized reality system's details from a historical perspective

/pdf/virtualized-reality-perspectives-on-4d-digitization-of-4sif2sac4c.pdf

Virtualized Reality: Perspectives on 4D Digitization of Dynamic Events

: In recent years, Binford's generalized cylinders have become an important tool for imagine understanding. However, research has been hampered by a lack of analytical results for these shapes. In this paper, a definition is presented for Straight Homogeneous Generalized Cylinders, those generalized cylinders, with a straight axis and with cross-sections which have constant shape but vary in size. This class of shapes, while still quite large, has properties which make considerable analysis possible. The results begin with deriving formulae for points and surface normals for these shapes. Theorems are presented concerning the conditions under which multiple descriptions can exist for a single solid shape. Then projections and contour generators are analyzed for some subclasses of shapes. The strongest results are obtained for solids of revolution (which the authors name Right Circular SHGCs), for which a closed-form method for analyzing image contours is presented. It is seen that a picture of the contours of a solid of revolution is ambiguous, with one degree of freedom related to the angle between the line of sight and the solid's axis. (Author)

The Theory of Straight Homogeneous Generalized Cylinders

We develop a computer vision-based system to transfer human motion from one subject to another. Our system uses a network of eight calibrated and synchronized cameras. We first build detailed kinematic models of the subjects based on our algorithms for extracting shape from silhouette across time (G. Cheung et al., 2003). These models are then used to capture the motion (joint angles) of the subjects in new video sequences. Finally we describe an image-based rendering algorithm to render the captured motion applied to the articulated model of another person. Our rendering algorithm uses an ensemble of spatially and temporally distributed images to generate photo-realistic video of the transferred motion. We demonstrate the performance of the system by rendering throwing and kungfu motions on subjects who did not perform them.

Takeo Kanade

Papers

Geometrical aspects of interpreting images as a three-dimensional scene

Gauge fixing for accurate 3D estimation

Virtualized Reality: Perspectives on 4D Digitization of Dynamic Events

The Theory of Straight Homogeneous Generalized Cylinders

Markerless human motion transfer