Donghoon Lee

Face alignment using cascade Gaussian process regression trees

In this paper, we propose a face alignment method that uses cascade Gaussian process regression trees (cGPRT) constructed by combining Gaussian process regression trees (GPRT) in a cascade stage-wise manner. Here, GPRT is a Gaussian process with a kernel defined by a set of trees. The kernel measures the similarity between two inputs as the number of trees where the two inputs fall in the same leaves. Without increasing prediction time, the prediction of cGPRT can be performed in the same framework as the cascade regression trees (CRT) but with better generalization. Features for GPRT are designed using shape-indexed difference of Gaussian (DoG) filter responses sampled from local retinal patterns to increase stability and to attain robustness against geometric variances. Compared with the previous CRT-based face alignment methods that have shown state-of-the-art performances, cGPRT using shape-indexed DoG features performed best on the HELEN and 300-W datasets which are the most challenging dataset today.

Face attribute classification using attribute-aware correlation map and gated convolutional neural networks.

This paper proposes a face attribute classification method based on attribute-aware correlation map and gated convolutional neural networks (CNN). The attribute-aware correlation map provides correlation information between pixel-location and attribute label, and each correlation map of an attribute provides information regarding regions where the relevant features should be extracted. Using the correlation maps of all the attributes, a number of most relevant face part regions are discovered. Based on the face part regions, gated columns of CNNs are simultaneously pre-trained on for face representations then fine-tuned for attribute classification. Here, each CNN column takes input from one of the regions discovered. The column of the CNN is gated such that in the backpropagation of the learning process, classification error due to less relevant attributes do not over influence the learning process. In the experiment, we manually labeled each image in the Labeled Faces in the Wild (LFW) benchmark dataset with 40 face attributes and obtained significant performance improvement over other state-of-the art methods.

A study on the rear-end collision warning system by considering different perception-reaction time using multi-layer perceptron neural network

A rear-end Collision Warning System (CWS) is applied for mitigating collision risk to the frontal motor vehicle under the traffic conditions. Most of the previous studies have been performed to address the braking behavior related problems based on the deterministic or stochastic parametric methods. However, these algorithms are of doubtful validity in the context of individual driving characteristics such as Perception-Reaction Time (PRT). This paper proposes a framework on Rear-end CWS to take into consideration of PRT effects based on the Artificial Neural Network (ANN). Multi-layer perceptron neural network based rear-end collision warning algorithm (MCWA) is developed and evaluated through a comparison between the conventional algorithms such as Time To Collision (TTC) and Stopping Distance Algorithm (SDA). The comparison study demonstrates that the proposed algorithm outperforms other traditional algorithms for detecting and predicting the rear-end collision risks. The proposed algorithm could be used for rear-end collision warning in car-following case without the influence of different human PRT.

Channel gain cartography via low rank and sparsity

Channel gain cartography aims at inferring the channel gains between arbitrary points in space based on measurements (samples) of channel gains taken from finite pairs of transceivers. Channel gain maps are useful for various sensing and resource allocation tasks, essential for the operation of cognitive radio networks. In this work, the channel gain samples are modeled as compressive tomographic measurements of an underlying spatial loss field (SLF), postulated to have low-rank structure corrupted by sparse errors. Efficient algorithms to reconstruct the SLF are developed, from which arbitrary channel gains can be interpolated.

Channel Gain Cartography for Cognitive Radios Leveraging Low Rank and Sparsity

Channel gain cartography aims at inferring the channel gains between two arbitrary points in space based on the measurements (samples) of the gains collected by a set of radios deployed in the area. Channel gain maps are useful for various sensing and resource allocation tasks essential for the operation of cognitive radio networks. In this paper, the channel gains are modeled as the tomographic accumulations of an underlying spatial loss field (SLF), which captures the attenuation in the signal strength due to the obstacles in the propagation path. In order to estimate the map accurately with a relatively small number of measurements, the SLF is postulated to have a low-rank structure possibly with sparse deviations. Efficient batch and online algorithms are derived for the resulting map reconstruction problem. Comprehensive tests with both synthetic and real data sets corroborate that the algorithms can accurately reveal the structure of the propagation medium, and produce the desired channel gain maps.

Salient object detection using bipartite dictionary

This paper considers a bipartite dictionary based salient object detection algorithm that assigns one of two labels (object/background) to each superpixel of an image. The algorithm will iteratively find for each of the labels two dictionaries referred to as the bipartite dictionary, and the dictionaries will in turn update the labels of the superpixels based on the assumption that features of a particular label is better represented by the dictionary of its own label than by the dictionary of the other label. This iteration stops when convergence is reached, in other words, when there is no update. An objective function is formulated such that the bipartite dictionary and superpixel labels maximize inter-class reconstruction error while simultaneously minimize intra-class reconstruction error. The proposed algorithm is evaluated on the MSRA-1000 dataset. Experimental results show that the proposed algorithm performs better than state-of-the-art algorithms for the dataset when the initial conditions are set appropriately. We have also found that the proposed algorithm tends to highlight salient objects more uniformly than other algorithms.

Joint Estimation of Pose and Face Landmark

This paper proposes a parallel joint boosting method that simultaneously estimates poses and face landmarks. The proposed method iteratively updates the poses and face landmarks through a cascade of parallel random ferns in a forward stage-wise manner. At each stage, the pose and face landmark estimates are updated: pose probabilities are updated based on previous face landmark estimates and face landmark estimates are updated based on previous pose probabilities. Both poses and face landmarks are simultaneously estimated through sharing parallel random ferns for the pose and face landmark estimations. This paper also proposes a triangular-indexed feature that references a pixel as a linear weighted sum of three chosen landmarks. This provides robustness against variations in scale, transition, and rotation. Compared with previous boosting methods, the proposed method reduces the face landmark error by 7.1 % and 12.3 % in the LFW and MultiPIE datasets, respectively, while it also achieves pose estimation accuracies of 78.6 % and 94.0 % in these datasets.

A transmission period selection scheme for device-to-device communications

In this paper, we propose a new power optimization and transmission period selection scheme for device-to-device (D2D) communication as an underlay to cellular networks. We show that the proposed scheme can achieve performance gain by flexibly allocating transmission power based on channel state information (CSI) and significantly increase reliability by selecting a transmission period.

Outage Analysis and Optimal Power allocation for Network-coding-based Hybrid AF and DF

Network coding was proposed to increase the achievable throughput of multicast in a network. Recently, combining network coding into user cooperation has attracted research attention. For cooperative transmission schemes with network coding, users combine their own and their partners messages by network coding. In previous works, it was shown that adaptive DF with network coding can achieve diversity gain and additional throughput gain. In this paper, to improve performance of conventional protocols and maximize advantage of using network coding, we propose a new network coding based user cooperation scheme which uses adaptively amplify-and-forward and decode-and-forward according to interuser channel status. We derive outage probability bound of proposed scheme and prove that it has full diversity order in the high SNR regime. Moreover, based on the outage bound, we compute optimal power allocation for the proposed scheme.

Learning based Utility Maximization for Multi-Resource Management: Poster Abstract

This poster addresses the problem of Network Utility Maximization (NUM) where multiple resources (computing/networking) participate in user services. NUM has usually been solved by Backpressure algorithms which has to build up queue size gradualy. This disadvantage stands out in the situation of multi-resource environment or multi-hop networking. To address the problem, we propose a reinforcement learning based algorithm that utilizes future prediction to overcome the previous limitation of non-learning based algorithms.