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

Current approaches to automated analysis have focused an a small set of prototypic expressions (e.g. joy or anger). Prototypic expressions occur infrequently in everyday life, however, and emotion expression is far more varied. To capture the full range of emotion expression, automated discrimination of fine grained changes in facial expression is needed. We developed and implemented an optical flow based approach (feature point tracking) that is sensitive to subtle changes in facial expression. In image sequences from 100 young adults, action units and action unit combinations in the brow and mouth regions were selected for analysis if they occurred a minimum of 25 times in the image database. Selected facial features were automatically tracked using a hierarchical algorithm for estimating optical flow. Image sequences were randomly divided into training and test sets. Feature point tracking demonstrated high concurrent validity with human coding using the Facial Action Coding System (FACS).

/pdf/feature-point-tracking-by-optical-flow-discriminates-subtle-2sois9hfy8.pdf

Feature-point tracking by optical flow discriminates subtle differences in facial expression

The visual medium evolved from early paintings to the realistic paintings of the classical era to photographs. The medium of moving imagery started with motion pictures. Television and video recording advanced it to show action "live" or capture and playback later. In all of the above media, the view of the scene is determined at the transcription time, independent of the viewer. We have been developing a new visual medium called virtualized reality. It delays the selection of the viewing angle until view time, using techniques from computer vision and computer graphics. The visual event is captured using many cameras that cover the action from all sides. The 3D structure of the event, aligned with the pixels of the image, is computed for a few selected directions using a stereo technique. Triangulation and texture mapping enable the placement of a "soft-camera" to reconstruct the event from any new viewpoint. With a stereo-viewing system, virtualized reality allows a viewer to move freely in the scene, independent of the transcription angles used to record the scene. We describe the hardware and software setup in our "studio" to make virtualized reality movies. Examples are provided to demonstrate the effectiveness of the system.

/pdf/virtualized-reality-concepts-and-early-results-3efnbg0s0d.pdf

Virtualized reality: concepts and early results

The Distant Human Identification (DHID) system is a master-slave, real-time surveillance system designed to acquire biometric imagery of humans at distance. A stationary wide field of view master camera is used to monitor an environment at distance. When the master camera detects a moving person, a narrow field of view slave camera is commanded to turn to that direction, acquire the target human, and track them while recording zoomed-in images. These zoomed-in views provide meaningful biometric imagery of the distant humans, who are not recognizable in the master view. Based on the lenses we currently use, the system can detect and track moving people at distances up to 50 meters, within a 60° field of regard.

/pdf/a-master-slave-system-to-acquire-biometric-imagery-of-humans-4j2ge3b62p.pdf

A master-slave system to acquire biometric imagery of humans at distance

We describe a novel camera calibration algorithm for square, circle, and ring planar calibration patterns An iterative refinement approach is proposed that utilizes the parameters obtained from traditional calibration algorithms as initialization to perform undistortion and unprojection of calibration images to a canonical fronto-parallel plane This canonical plane is then used to localize the calibration pattern control points and recompute the camera parameters in an iterative refinement until convergence Undistorting and unprojecting the calibration pattern to the canonical plane increases the accuracy of control point localization and consequently of camera calibration We have conducted an extensive set of experiments with real and synthetic images for the square, circle and ring pattern, and the pixel reprojection errors obtained by our method are about 50% lower than those of the OpenCV Camera Calibration Toolbox Increased accuracy of camera calibration directly leads to improvements in other applications; we demonstrate recovery of fine object structure for visual hull reconstruction, and recovery of precise epipolar geometry for stereo camera calibration

/pdf/accurate-camera-calibration-using-iterative-refinement-of-2t045smpd9.pdf

Accurate camera calibration using iterative refinement of control points

In this paper a process is described for analysing the motion of a human target in a video stream. Moving targets are detected and their boundaries extracted. From these, a \"star\" skeleton is produced. Two motion cues are determined from this skeletonization: body posture, and cyclic motion of skeleton segments. These cues are used to determine human activities such as walking or running, and even potentially, the target's gait. Unlike other methods, this does not require an a priori human model, or a large number of \"pixels on target\". Furthermore, it is computationally inexpensive, and thus ideal for real-world video applications such as outdoor video surveillance.

Takeo Kanade

Papers

Feature-point tracking by optical flow discriminates subtle differences in facial expression

Virtualized reality: concepts and early results

A master-slave system to acquire biometric imagery of humans at distance

Accurate camera calibration using iterative refinement of control points

Real-Time Human Motion Analysis by Image Skeletonization