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

Detection, tracking, and classification of action units in facial expression

Scene flow is the three-dimensional motion field of points in the world, just as optical flow is the two-dimensional motion field of points in an image. Any optical flow is simply the projection of the scene flow onto the image plane of a camera. We present a framework for the computation of dense, non-rigid scene flow from optical flow. Our approach leads to straightforward linear algorithms and a classification of the task into three major scenarios: complete instantaneous knowledge of the scene structure; knowledge only of correspondence information; and no knowledge of the scene structure. We also show that multiple estimates of the normal flow cannot be used to estimate dense scene flow directly without some form of smoothing or regularization.

Three-dimensional scene flow

Computer analysis of human face images includes detection of faces, identification of people, and understanding of expression. Among these three tasks, facial expression has been the least studied, and most of the past work on facial expression tried to recognize a small set of emotions, such as joy, disgust, and surprise. This practice may follow from the work of Darwin, who proposed that emotions have corresponding prototypic facial expressions. In everyday life, however, such prototypic expressions occur relatively infrequently; instead, emotion is communicated more often by subtle changes in one or a few discrete features. FACS-code Action Units, defined by Ekman, are one such representation accepted in the psychology community.

In collaboration with psychologists, we have been developing a system for automatically recognizing facial action units. This talk will present the current version of the system. The system uses a 3D Active Appearance Model to align a face image and transform it to a person-specific canonical coordinate frame. This transformation can remove appearance changes due to changes of head pose and relative illumination direction. In this transformed image frame, we perform detailed analysis of both facial motion and facial appearance changes, results of which are fed to an action-unit recogniser.

Facial expression analysis

This thesis uses a physical, intrinsic reflection model to analyze the effects of shading and highlights on dielectric, non-uniform materials, such as plastics and paints. Since black-and-white images do not provide sufficient information to interpret these effects appropriately, this work uses color images. The colors seen on a dielectric object are an additive mixture of the object color and the illumination color, i.e.: the colors are a linear combination of those two color vectors. They therefore lie in a plane in the color space. Within the plane, they form a cluster that looks like a skewed T, due to the geometric properties of shading and highlight reflection.
When combined with a camera model, this description of physically possible color changes on one object can be used to guide an algorithm in interpreting color images. It is the basis of an algorithm that analyzes color variation in real color images. The algorithm looks for characteristic color clusters in the image and relates their shape to hypotheses on the object color and the amount of shading, as well as the position, strength and color of highlights on the object. It verifies the hypotheses by reapplying them back to the image, testing whether and how they apply to the pixels in the local image area from which the respective hypothesis was derived. In this way, it adapts the image interpretation process to local scene characteristics and reacts differently to color and intensity changes at different places in the image.
The result is an image segmentation that ignores color changes due to shading and highlights, as well as a set of intrinsic images, one describing the amount of shading at every pixel in the image, and the other one showing the amount highlight reflection at every pixel. Such intrinsic images can be a useful tool for other areas of computer vision that are currently disturbed by highlights in images, such as stereo vision and motion analysis. The approach can also be used to determine the illumination color from the highlights and as a preprocessor for methods to determine object shape from shading or highlights. This line of research may lead to physics-based image understanding methods that are both more reliable and more useful than traditional methods. (Abstract shortened with permission of author.)

/pdf/a-physical-approach-to-color-image-understanding-34sisikqri.pdf

A physical approach to color image understanding

There has been a recent push in extraction of 3D spatial layout of scenes. However, none of these approaches model the 3D interaction between objects and the spatial layout. In this paper, we argue for a parametric representation of objects in 3D, which allows us to incorporate volumetric constraints of the physical world. We show that augmenting current structured prediction techniques with volumetric reasoning significantly improves the performance of the state-of-the-art.

/pdf/estimating-spatial-layout-of-rooms-using-volumetric-3gzkacxfwy.pdf

Takeo Kanade

Papers

Detection, tracking, and classification of action units in facial expression

Three-dimensional scene flow

Facial expression analysis

A physical approach to color image understanding

Estimating Spatial Layout of Rooms using Volumetric Reasoning about Objects and Surfaces