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

This paper is concerned with the use of a database to support automated photo interpretation. The function of the database is to provide an environment in which to perform photo interpretation utilizing software tools, and represent domain knowledge about the scenes being interpreted. Within the framework of the database, image interpretation systems use knowledge stored as map, terrain, or scene descriptions to provide structural or spatial constraints to guide human and machine processing. We describe one such system under development, MAPS (Map Assisted Photo interpretation System), and give some general rationales for its design and implementation.© (1981) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Data Base Support For Automated Photo Interpretation

Object representation is useful for many computer vision tasks, such as object detection, recognition, and tracking. Computer vision tasks must handle situations where unknown objects appear and must detect and track some object which is not in the trained database. In such cases, the system must learn or, otherwise derive, descriptions of new objects. In this paper, we investigate creating a representation of previously unknown objects that newly appear in the scene. The representation creates a viewpoint-invariant and scale-normalized model approximately describing an unknown object with multimodal sensors. Those properties of the representation facilitate 3D tracking of the object using 2D-to-2D image matching. The representation has both benefits of an implicit model (referred to as a view-based model) and an explicit model (referred to as a shape-based model). Experimental results demonstrate the viability of the proposed representation and outperform the existing approaches for 3D-pose estimation.

Online approximate model representation of unknown objects

This thesis presents a method for the 3D digitization of dynamic, real-world events. This task requires sufficient temporal and spatial sampling to capture the entirety of an event, as well as the estimation of 3D shape and appearance over time. Direct sensing of global 3D structure is not feasible because of the motion of the scene, and even range scanning systems usually sample too coarsely or too slowly to accurately capture dynamic events. This thesis presents a method of 3D digitization that overcomes this sensing problem through the use of a synchronized collection of a large number of calibrated video cameras.
Our 3D digitization method decomposes 3D shape recovery into the estimation of visible structure in each video frame followed by the integration of visible structure into a complete 3D model. Visible surfaces are extracted using the multi-baseline stereo (MBS) algorithm. This implementation of MBS efficiently supports any number cameras in general positions through a novel rectification strategy for general camera configurations that do not allow the rectification of all images to a single 3D plane. Stereo-computed range images are then integrated within a volumetric space using a novel integration strategy. Each range image is converted into a 3D mesh, and then into an implicit surface embedded in the volumetric space by encoding the signed distance between each 3D volume sample (voxel) and the 3D mesh. Multiple range images are integrated by accumulating the signed distance at each voxel. The resulting global surface model is then extracted by applying the Marching Cubes implicit-surface extraction algorithm.
The digitization of scene appearance uses the estimated 3D structure to determine visibility and sampling of color in the original video images. A compact, global texture map is computed by mixing the color estimates, emphasizing the color estimates obtained from cameras viewing the local surface structure most directly. Alternatively, an image-based representation is derived from the global model by reprojecting the global structure itself back into the original cameras to generate a visible surface model in each real camera. New views are synthesized by separately rendering three of these models and mixing the rendered views of the models directly on the virtual image plane.
This thesis also presents extensive results of digitizing real events recorded in the 3D Dome, a recording studio employing a synchronized collection of 51 calibrated video cameras mounted on a 5-meter diameter geodesic dome. This facility provides a workspace of approximately 8 cubic meters, sufficient to capture the entirety of motion of 1-2 people performing athletic actions including swinging a baseball bat, bumping a volleyball, and passing a basketball. Results are presented for 5 different events, each 1-2 seconds long. Video is captured at 240x320 image resolution, with a volumetric modeling resolution of 1 cubic center. The resulting models are used to generate both simple camera motions near the original camera viewpoints as well as camera motions deep into the event space and into views nearly impossible to capture with real cameras.

A multi-camera method for three-dimensional digitization of dynamic, real-world events

The parameters estimated by Structure from Motion (SFM) contain inherent indeterminacies which we call gauge freedoms. Under a perspective camera, shape and motion parameters are only recovered up to an unknown similarity transformation. In this paper we investigate how covariance-based uncertainty can be represented under these gauge freedoms. Past work on uncertainty modeling has implicitly imposed gauge constraints on the solution before considering covariance estimation. Here we examine the effect of selecting a particular gauge on the uncertainty of parameters. We show potentially dramatic effects of gauge choice on parameter uncertainties. However the inherent geometric uncertainty remains the same irrespective of gauge choice. We derive a Geometric Equivalence Relationship with which covariances under different parametrizations and gauges can be compared, based on their true geometric uncertainty. We show that the uncertainty of gauge invariants exactly captures the geometric uncertainty of the solution, and hence provides useful measures for evaluating the uncertainty of the solution. Finally we propose a fast method for covariance estimation and show its correctness using the Geometric Equivalence Relationship.

Uncertainty modeling for optimal structure from motion

We present a novel feature representation for categorical object detection. Unlike previous approaches that have concentrated on generic interest-point detectors, we construct object-specific features directly from the training images. Our feature is represented by a collection of Flexible Edge Arrangement Templates (FEATs). We propose a two-stage semi-supervised learning approach to feature selection. A subset of frequent templates are first selected from a large template pool. In the second stage, we formulate feature selection as a regression problem and use LASSO method to find the most discriminative templates from the preselected ones. FEATs adaptively capture the image structure and naturally accommodate local shape variations. We show that this feature can be complemented by the traditional holistic patch method, thus achieving both efficiency and accuracy. We evaluate our method on three well-known car datasets, showing performance competitive with existing methods.

Takeo Kanade

Papers

Data Base Support For Automated Photo Interpretation

Online approximate model representation of unknown objects

A multi-camera method for three-dimensional digitization of dynamic, real-world events

Uncertainty modeling for optimal structure from motion

Flexible Edge Arrangement Templates for Object Detection