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

A VLSI sensor based rangefinding system

This paper describes a method for detecting multiple overlapping objects from a real-time video stream Layered detection is based on two processes: pixel analysis and region analysis Pixel analysis determines whether a pixel is stationary or transient by observing its intensity over time Region analysis detects regions consisting of stationary pixels corresponding to stopped objects These regions are registered as layers on the background image, and thus new moving objects passing through these layers can be detected An important aspect of this work derives from the observation that legitimately moving objects in a scene tend to cause much faster intensity transitions than changes due to lighting, meteorological, and diurnal effects The resulting system robustly detects objects at an outdoor surveillance site For 8 hours of video evaluation, a detection rate of 92% was measured which is higher than traditional background subtraction methods

/pdf/layered-detection-for-multiple-overlapping-objects-u8um5212pm.pdf

Layered detection for multiple overlapping objects

Subspace clustering has many applications in computer vision, such as image/video segmentation and pattern classification. The major issue in subspace clustering is to obtain the most appropriate subspace from the given noisy data. Typical methods (e.g., SVD, PCA, and eigen-decomposition) use least squares techniques, and are sensitive to outliers. In this paper, we present the k-th nearest neighbor distance (kNND) metric, which, without actually clustering the data, can exploit the intrinsic data cluster structure to detect and remove influential outliers as well as small data clusters. The remaining data provide a good initial inlier data set that resides in a linear subspace whose rank (dimension) is upper-bounded. Such linear subspace constraint can then be exploited by simple algorithms, such as iterative SVD algorithm, to (1) detect the remaining outliers that violate the correlation structure enforced by the low rank subspace, and (2) reliably compute the subspace. As an example, we apply our method to extracting layers from image sequences containing dynamically moving objects.

/pdf/robust-subspace-clustering-by-combined-use-of-knnd-metric-4r1042j2n9.pdf

Robust subspace clustering by combined use of kNND metric and SVD algorithm

This edited and reviewed volume consists of papers that were originally presented at a workshop in the Scientific Center at Schloss Dagstuhl, Germany. It gives an overview of the field and presents the latest developments in the areas of modelling and planning for sensor based robots. The particular topics addressed include active vision, sensor fusion, environment modelling, motion planning, robot navigation, distributed control architectures, reactive behaviour, and others. Key contributors to this volume are: S. Arimoto, R. Bolles, H. Bunke, H. Christensen, R. Dillmann, T. Henderson, R. Jarvis, H. Noltemeier, U. Rembold, Y. Shirai, C. Torras.

Modelling and Planning for Sensor Based Intelligent Robot Systems

Research in early (low-level) vision, tooth for machines and humans, has
traditionally been based on the study of idealized images or image patches such as step edges, gratings, flat fields, and Mondrians. Real images, however, exhibit much richer and more complex structure, whose nature is determined by the physical and geometric properties of illumination, reflection, and imaging. By understanding these physical relationships, a new kind of early vision analysis is made possible. In this paper, we describe a progression of models of imaging physics that present a much more complex and realistic set of image relationships than are commonly assumed in early vision research. We begin with the Dichromatic Reflection Model, which describes how highlights and color are related in images of dielectrics such as plastic and painted surfaces. This gives rise to a mathematical relationship in color space to separate highlights from object color. Perceptions of shape, surface roughness/texture, and illumination color are readily derived from this analysis. We next show how this can be extended to images of several objects, by deriving local color variation relationships from the basic model. The resulting method for color image analysis has been successfully applied in machine vision experiments in our laboratory. Yet another extension is to account for inter-reflection among multiple objects.
We have derived a simple model of color inter-reflection that accounts for the basic phenomena, and report on this model and how we are applying it. In general, the concept of illumination for vision should account for the entire "illumination environment", rather than being restricted to a single light source. This work shows that the basic physical relationships give rise to very structured image properties, which can be a more valid basis for early vision than the traditional idealized image patterns.

Takeo Kanade

Papers

A VLSI sensor based rangefinding system

Layered detection for multiple overlapping objects

Robust subspace clustering by combined use of kNND metric and SVD algorithm

Modelling and Planning for Sensor Based Intelligent Robot Systems

Physics-based models for early vision by machine