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

The authors are prototyping a legged vehicle, the Ambler, for an exploratory mission on another planet, conceivably Mars, where it is to traverse uncharted areas and collect material samples. They describe how the rover can construct from range imagery a geometric terrain representation, i.e., elevation map that includes uncertainty, unknown areas, and local features. First, they present an algorithm for constructing an elevation map from a single range image. By virtue of working in spherical-polar space, the algorithm is independent of the desired map resolution and the orientation of the sensor, unlike algorithms that work in Cartesian space. Secondly, the authors present a two-stage matching technique (feature matching followed by iconic matching) that identifies the transformation T corresponding to the vehicle displacement between two viewing positions. Thirdly, to support legged locomotion over rough terrain, they describe methods for evaluating regions of the constructed elevation maps as footholds. >

/pdf/terrain-mapping-for-a-roving-planetary-explorer-3dm415pn8q.pdf

Terrain mapping for a roving planetary explorer

Recovery of three dimensional (3D) shape and motion of non-static scenes from a monocular video sequence is important for applications like robot navigation and human computer interaction. If every point in the scene randomly moves, it is impossible to recover the non-rigid shapes. In practice, many non-rigid objects, e.g. the human face under various expressions, deform with certain structures. Their shapes can be regarded as a weighted combination of certain shape bases. Shape and motion recovery under such situations has attracted much interest. Previous work on this problem (Bregler, C., Hertzmann, A., and Biermann, H. 2000. In Proc. Int. Conf. Computer Vision and Pattern Recognition; Brand, M. 2001. In Proc. Int. Conf. Computer Vision and Pattern Recognition; Torresani, L., Yang, D., Alexander, G., and Bregler, C. 2001. In Proc. Int. Conf. Computer Vision and Pattern Recognition) utilized only orthonormality constraints on the camera rotations (rotation constraints). This paper proves that using only the rotation constraints results in ambiguous and invalid solutions. The ambiguity arises from the fact that the shape bases are not unique. An arbitrary linear transformation of the bases produces another set of eligible bases. To eliminate the ambiguity, we propose a set of novel constraints, basis constraints, which uniquely determine the shape bases. We prove that, under the weak-perspective projection model, enforcing both the basis and the rotation constraints leads to a closed-form solution to the problem of non-rigid shape and motion recovery. The accuracy and robustness of our closed-form solution is evaluated quantitatively on synthetic data and qualitatively on real video sequences.

/pdf/a-closed-form-solution-to-non-rigid-shape-and-motion-eld3q9md68.pdf

A Closed-Form Solution to Non-Rigid Shape and Motion Recovery

Charge-Coupled Device (CCD) cameras are widely used imaging sensors in computer vision systems. Many photometric algorithms, such as shape from shading, color constancy and photometric stereo, implicitly assume that the image intensity is proportional to scene radiance. The actual image measurements deviate significantly from this assumption since the transformation from scene radiance to image intensity is non-linear and is a function of various factors including: noise sources in the CCD sensor, as well as various transformations occurring in the camera including: white balancing, gamma correction and automatic gain control. This paper illustrates how careful modeling of the error sources and the various processing steps enable us to accurately estimate the "response function", the inverse mapping from image measurements to scene radiance for a given camera exposure setting. It is shown that the estimation algorithm outperforms the calibration procedures known to us in terms of reduced bias and variance. Further, we demonstrate how the error modelling helps us to obtain uncertainty estimates of the camera irradiance value. The power of this uncertainty modeling is illustrated by a vision task involving High Dynamic Range image generation followed by change detection. Change can be detected reliably even in situation where the two images (the reference scene image and the current image) are taken several hours apart.

/pdf/statistical-calibration-of-ccd-imaging-process-1t313j4ns5.pdf

Statistical calibration of CCD imaging process

A method of virtualizing reality, i.e., a method of creating virtual reality from images of a real event, is comprised of the steps of capturing a plurality of images of each time instant of a real event using a plurality of cameras positioned at a plurality of angles. Each image is stored as intensity and/or color information. A suitable internal representation is computed from these images and the information regarding the camera angles. An image of each time instant may be generated from any viewing angle using the internal representation of it. The virtual viewpoints could be displayed on a single TV screen or using a stereoscopic display device for a true three-dimensional effect. The event thus virtualized can be navigated through, and interacted with, any virtual reality system.

Method for creating virtual reality

The structure from motion problem has been extensively studied in the field of computer vision. Yet, the bulk of the existing work assumes that the scene contains only a single moving object. The more realistic case where an unknown number of objects move in the scene has received little attention, especially for its theoretical treatment. We present a new method for separating and recovering the motion and shape of multiple independently moving objects in a sequence of images. The method does not require prior knowledge of the number of objects, nor is dependent on any grouping of features into an object at the image level. For this purpose, we introduce a mathematical construct of object shapes, called the shape interaction matrix, which is invariant to both the object motions and the selection of coordinate systems. This invariant structure is computable solely from the observed trajectories of image features without grouping them into individual objects. Once the structure is computed, it allows for segmenting features into objects by the process of transforming it into a canonical form, as well as recovering the shape and motion of each object. >

Takeo Kanade

Papers

Terrain mapping for a roving planetary explorer

A Closed-Form Solution to Non-Rigid Shape and Motion Recovery

Statistical calibration of CCD imaging process

Method for creating virtual reality

A multi-body factorization method for motion analysis