This paper describes a sophisticated method to estimate visual direction using iris contours. This method requires only one monocular image taken by a camera with unknown focal length. In order to estimate the visual direction, we assume the visual directions of both eyes are parallel and iris boundaries are circles in 3D space. In this case, the two planes where the iris boundaries reside are also parallel. We estimate the normal vector of the two planes from the iris contours extracted from an input image by using an extended "two-circle" algorithm. Unlike most existing gaze estimation algorithms that require information about eye corners and heuristic knowledge about 3D structure of the eye in addition to the iris contours, our method uses two iris contours only. Another contribution of our method is the ability of estimating the focal length of the camera. It allows one to use a zoom lens to take images and the focal length can be adjusted at any time. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This paper presents a novel algorithm for Nearest Neighbor (NN) classifier. NN classification is a well-known method of pattern classification having the following properties: * it performs maximum-margin classification and achieves less than twice the ideal Bayesian error, * it does not require knowledge of pattern distributions, kernel functions or base classifiers, and * it can naturally be applied to multiclass classification problems. Among the drawbacks are A) inefficient memory use and B) ineffective pattern classification speed. This paper deals with the problems A and B. In most cases, NN search algorithms, such as k-d tree, are employed as a pattern search engine of the NN classifier. However, NN classification does not always require the NN search. Based on this idea, we propose a novel algorithm named k-d decision tree (KDDT). Since KDDT uses Voronoi-condensed prototypes, it consumes less memory than naive NN classifiers. We have confirmed that KDDT is much faster than NN search-based classifier through a comparative experiment (from 9 to 369 times faster than NN search based classifier). Furthermore, in order to extend applicability of the KDDT algorithm to high-dimensional NN classification, we modified it by incorporating Gabriel editing or RNG editing instead of Voronoi condensing. Through experiments using simulated and real data, we have confirmed the modified KDDT algorithms are superior to the original one.

Pruning is an effective technique to reduce computational complexity of Convolutional Neural Networks (CNNs) by removing redundant neurons (or weights). There are two types of pruning methods: holistic pruning and layer-wise pruning. The former selects the least important neuron from the entire model and prunes it. The latter conducts pruning layer by layer. Recently, it has turned out that some layer-wise methods are effective for reducing computational complexity of pruned models while preserving their accuracy. The difficulty of layer-wise pruning is how to adjust pruning ratio (the ratio of neurons to be pruned) in each layer. Because CNNs typically have lots of layers composed of lots of neurons, it is inefficient to tune pruning ratios by human hands. In this paper, we present Pruning Ratio Optimizer (PRO), a method that can be combined with layer-wise pruning methods for optimizing pruning ratios. The idea of PRO is to adjust pruning ratios based on how much pruning in each layer has an impact on the outputs in the final layer. In the experiments, we could verify the effectiveness of PRO.

This paper presents an RK-means clustering algorithm which is developed for reliable data grouping by introducing a new reliability evaluation to the K-means clustering algorithm. The conventional K-means clustering algorithm has two shortfalls: 1) the clustering result will become unreliable if the assumed number of the clusters is incorrect; 2) during the update of a cluster center, all the data points belong to that cluster are used equally without considering how distant they are to the cluster center. In this paper, we introduce a new reliability evaluation to K-means clustering algorithm by considering the triangular relationship among each data point and its two nearest cluster centers. We applied the proposed algorithm to track objects in video sequence and confirmed its effectiveness and advantages.

In this paper, we present a general object tracking method based on a newly proposed pixel-wise clustering algorithm. To track an object in a cluttered environment is a challenging issue because a target object may be in concave shape or have apertures (e.g. a hand or a comb). In those cases, it is difficult to separate the target from the background completely by simply modifying the shape of the search area. Our algorithm solves the problem by 1) describing the target object by a set of pixels; 2) using a K-means based algorithm to detect all target pixels. To realize stable and reliable detection of target pixels, we firstly use a 5D feature vector to describe both the color ("Y, U, V") and the position ("x, y") of each pixel uniformly. This enables the simultaneous adaptation to both the color and geometric features during tracking. Secondly, we use a variable ellipse model to describe the shape of the search area and to model the surrounding background. This guarantees the stable object tracking under various geometric transformations. The robust tracking is realized by classifying the pixels within the search area into "target" and "background" groups with a K-means clustering based algorithm that uses the "positive" and "negative" samples. We also propose a method that can detect the tracking failure and recover from it during tracking by making use of both the "positive" and "negative" samples. This feature makes our method become a more reliable tracking algorithm because it can discover the target once again when the target has become lost. Through the extensive experiments under various environments and conditions, the effectiveness and efficiency of the proposed algorithm is confirmed.

This paper presents a pruning method, Reconstruction Error Aware Pruning (REAP), to reduce the redundancy of convolutional neural network models for accelerating their inference. In REAP, we have the following steps: 1) Prune the channels whose outputs are redundant and can be reconstructed from the outputs of other channels in each convolutional layer; 2) Update the weights of the remaining channels by least squares method so as to compensate the error caused by pruning. This is how we compress and accelerate the models that are initially large and slow with little degradation. The ability of REAP to maintain the model performances saves us lots of time and labors for retraining the pruned models. The challenge of REAP is the computational cost for selecting the channels to be pruned. For selecting the channels, we need to solve a huge number of least squares problems. We have developed an efficient algorithm based on biorthogonal system to obtain the solutions of those least squares problems. In the experiments, we show that REAP can conduct pruning with smaller sacrifice of the model performances than several existing methods including the previously state-of-the-art one.

This paper presents a method for reducing the redundancy in both fully connected layers and convolutional layers of trained neural network models. The proposed method consists of two steps, 1) Neuro-Coding: to encode the behavior of each neuron by a vector composed of its outputs corresponding to actual inputs and 2) Neuro-Unification: to unify the neurons having the similar behavioral vectors. Instead of just pruning one of the similar neurons, the proposed method let the remaining neuron emulate the behavior of the pruned one. Therefore, the proposed method can reduce the number of neurons with small sacrifice of accuracy without retraining. Our method can be applied for compressing convolutional layers as well. In the convolutional layers, the behavior of each channel is encoded by its output feature maps, and channels whose behaviors can be well emulated by other channels are pruned and update the remaining weights. Through several experiments, we comfirmed that the proposed method performs better than the existing methods.