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

2P1-N-128 Design of an MR-compatible three-axis force sensor(Medical and Welfare Robotics and Mechatronics 4,Mega-Integration in Robotics and Mechatronics to Assist Our Daily Lives)

: This project investigated technologies to enable autonomous flight of agile vehicles in urban environments. Specifically, technologies were developed that related to vision-based feedback for control. The mission profile under consideration was a single vehicle carrying a video camera while flying below the rooftops of a city with no additional sensors or pre-existing map information. As such, substantial progress was made in the areas of feature-point tracking, state estimation, scene reconstruction, robustness to camera calibration, daisy-chaining navigation, mapping, path planning, and feedback characterization. An integrated approach was used that (be used on multi-disciplinary analysis for decision making and control commands. The project spiraled the maturation of technologies from theory to simulation to flight testing. The simulation relied upon a hardware-in-the-loop simulation facility that allowed the physical camera to record measurements from high-fidelity graphics and thus consider the effects of distortion and nonlinearities.

Vision-Based Control of Agile, Autonomous Micro Air Vehicles and Small UAVs in Urban Environments

This report describes progress in research on an autonomous robot for planetary exploration. In 1989, the year covered by this report, a six-legged walking robot, the Ambler, was configured, designed, and constructed. This configuration was used to overcome shortcomings exhibited by existing wheeled and walking robot mechanisms. The fundamental advantage of the Ambler is that the actuators for body support are independent of those for propulsion; a subset of the planar joints propel the body, and the vertical actuators support and level the body over terrain. Models of the Ambler's dynamics were developed and the leveling control was studied. An integrated system capable of walking with a single leg over rugged terrain was implemented and tested. A prototype of an Ambler leg is suspended below a carriage that slides along rails. To walk, the system uses a laser scanner to find a clear, flat foothold, positions the leg above the foothold, contacts the terrain with the foot, and applies force enough to advance the carriage along the rails. Walking both forward and backward, the system has traversed hundreds of meters of rugged terrain including obstacles too tall to step over, trenches too deep to step in, closely spaced rocks, and sand hills. In addition, preliminary experiments were conducted with concurrent planning and execution, and a leg recovery planner that generates time and power efficient 3D trajectories using 2D search was developed. A Hero robot was used to demonstrate mobile manipulation. Indoor tasks include collecting cups from the lab floor, retrieving printer output, and recharging when its battery gets low. The robot monitors its environment, and handles exceptional conditions in a robust fashion, using vision to track the appearance and disappearance of cups, onboard sonars to detect imminent collisions, and monitors to detect the battery level.

/pdf/autonomous-planetary-rover-at-carnegie-mellon-2os4m191p3.pdf

Autonomous Planetary Rover at Carnegie Mellon

Angiography is a method used by radiologists to examine the structure and health of blood vessels. The method consists of injecting an x-ray opaque contrast agent into the bloodstream, and taking one or more x-ray images of the vessel. Although these images are clear and have high resolution, they individually show structures in only two-dimensions. The goal of our work is to automatically reconstruct the threedimensional structure of a blood vessel using a small number of angiogram images taken from different angles. Among the types of information about blood vessels that are useful to physicians, the most important for diagnostic purposes is a measure of the cross-sectional area of the vessel at every location along its length. This information reveals constrictions caused by blood clots, cholesterol build-ups, or injuries. Treatment of these conditions may involve inserting a catheter, with a small video camera and scraping tool mounted at its tip, into the vessel to remove the obstruction. A computer can aid in planning this surgery by simulating what the tip of the catheter would encounter in the vessel, and by computing how far the catheter must be inserted to begin repair. This paper presents an algorithm which performs the three-dimensional reconstruction task. Each section of the paper will examine one of the main steps of the algorithm: detecting vessels in single images, finding the positions of the vessels in three dimensions, and finally performing diameter measurements. The paper concludes with some reconstruction results.

/pdf/reconstructing-3-d-blood-vessel-shapes-from-multiple-x-ray-44clsjwctl.pdf

Reconstructing 3-D Blood Vessel Shapes from Multiple X-Ray Images

Face alignment is the problem of automatically locating detailed facial landmarks across different subjects, illuminations, and viewpoints. Previous methods can be divided into two broad categories. 2D-based methods locate a relatively small number of 2D fiducial points in real time while 3D-based methods fit a high-resolution 3D model offline at a much higher computational cost.

Takeo Kanade

Papers

2P1-N-128 Design of an MR-compatible three-axis force sensor(Medical and Welfare Robotics and Mechatronics 4,Mega-Integration in Robotics and Mechatronics to Assist Our Daily Lives)

Vision-Based Control of Agile, Autonomous Micro Air Vehicles and Small UAVs in Urban Environments

Autonomous Planetary Rover at Carnegie Mellon

Reconstructing 3-D Blood Vessel Shapes from Multiple X-Ray Images

Real-time dense 3D face alignment from 2D video with automatic facial action unit coding