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

A two-step procedure that finds natural clusters in geometric point data is described. The first step computes a hierarchical cluster tree minimizing an entropy objective function. The second step recursively explores the tree for a level clustering having minimum description length. Together, these two steps find natural clusters without requiring a user to specify threshold parameters or so-called magic numbers. In particular, the method automatically determines the number of clusters in the input data. The first step exploits a new hierarchical clustering procedure called numerical iterative hierarchical clustering (NIHC). The output of NIHC is a cluster tree. The second step in the procedure searches the tree for a minimum-description-length (MDL) level clustering. The MDL formulation, equivalent to maximizing the posterior probability, is suited to the clustering problem because it defines a natural prior distribution. >

Finding natural clusters having minimum description length

The complete set of measurements that could ever be used by a stereo algorithm is the plenoptic function or light-field. We give a concise characterization of when the light-field of a Lambertian scene uniquely determines its shape, and, conversely, when stereo is inherently ambiguous. We show that stereo computed from the complete light-field is ambiguous if and only if the scene is radiating light of a constant intensity (and color) over an extended region.

/pdf/a-characterization-of-inherent-stereo-ambiguities-22sepaiya2.pdf

A characterization of inherent stereo ambiguities

Super-resolution is usually posed as a reconstruction problem. The low resolution input images are assumed to be noisy, downsampled versions of an unknown super-resolution image that is to be estimated. A common way of inverting the down-sampling process is to write down the reconstruction constraints and then solve them, often adding a smoothness prior to regularize the solution. In this paper, we present two results which both show that there is more to super-resolution than image reconstruction. We first analyze the reconstruction constraints and show that they provide less and less useful information as the magnification factor increases. Afterwards, we describe a “hallucination” algorithm, incorporating the recognition of local features in the low resolution images, which outperforms existing reconstruction-based algorithms.

Super-resolution: reconstruction or recognition?

Object discovery algorithms group together image regions that originate from the same object. This process is effective when the input collection of images contains a large number of densely sampled views of each object, thereby creating strong connections between nearby views. However, existing approaches are less effective when the input data only provide sparse coverage of object views.

We propose an approach for object discovery that addresses this problem. We collect a database of about 5 million product images that capture 1.2 million objects from multiple views. We represent each region in the input image by a "bag" of database object regions. We group input regions together if they share similar "bags of regions." Our approach can correctly discover links between regions of the same object even if they are captured from dramatically different viewpoints. With the help from these added links, our proposed approach can robustly discover object instances even with sparse coverage of the viewpoints.

/pdf/connecting-missing-links-object-discovery-from-sparse-47za67167k.pdf

Connecting missing links: object discovery from sparse observations using 5 million product images

Computer sensing of hand and limb motion is an important problem for applications in human-computer interaction, virtual reality, and athletic performance measurement. We describe a framework for local tracking of self-occluding motion, in which parts of the mechanism obstruct each others visibility to the camera. Our approach uses a kinematic model to predict occlusion and windowed templates to track partially occluded objects. We analyze our model of self-occlusion, discuss the implementation of our algorithm, and give experimental results for 3D hand tracking under significant amounts of self-occlusion. These results extend the DigitEyes system for articulated tracking, which we have previously developed, to handle self-occluding motions.

http://www.ri.cmu.edu/pub_files/pub2/rehg_jim_1994_1/rehg_jim_1994_1.pdf

Takeo Kanade

Papers

Finding natural clusters having minimum description length

A characterization of inherent stereo ambiguities

Super-resolution: reconstruction or recognition?

Connecting missing links: object discovery from sparse observations using 5 million product images

Visual Tracking of Self-Occluding Articulated Objects